TLDR: The MIT NANDA initiative found that 95 per cent of enterprise generative AI pilots deliver no measurable impact on the profit and loss account, despite an estimated €27.6bn to €36.8bn of enterprise investment. This is not primarily a failure of the technology. Cost is legible because it is billed by the vendor; return is illegible because it is diffuse, delayed and rarely instrumented. This playbook sets out a complete operating system for closing that gap: a seven-band taxonomy of where AI spend actually goes, the financial equations that turn a hunch into a governance instrument, a seven-layer control structure, and the quantitative crossover that determines when owning your own models beats paying per unit of consumption.
Best For: CFOs, COOs, chief digital officers and transformation leads at Irish pharmaceutical, biotech and MedTech organisations who are past the pilot stage and need a defensible way to answer the question every board eventually asks: what did we get back?
Want the full white paper?
Get the complete 29-page PDF — every table, equation, figure and the full 26-source reference list, in one document.
Across boardrooms in Dublin, Cork and Limerick, the same sentence is being repeated in slightly different accents: we are spending on advanced computation, but we cannot yet prove the return. They are not imagining it. The MIT NANDA initiative, studying 300 public deployments alongside 150 executive interviews and 350 employee surveys, found that 95 per cent of enterprise generative pilots delivered no measurable impact on the profit and loss account, despite an estimated €27.6bn to €36.8bn of enterprise investment.1
This paper argues the shortfall is not, primarily, a failure of technology. It is a failure of governance. Cost is legible because it is billed. Return is illegible because it is diffuse, delayed and rarely instrumented. The organisations that will win the next five years are not the ones that spend the most, nor the ones that spend the least. They are the ones that govern the gap between a knowable cost and an unknowable return.
Currency note: figures below are stated in euro. Where a cited source reports in US dollars, the figure has been converted at €1 = US$1.09.
1. The ROI paradox: why cost is legible and return is not
The defining feature of enterprise AI adoption in 2026 is an asymmetry of legibility. The cost is precise to four decimal places. The return is a rumour.
Read any of the industry post-mortems and a pattern emerges. Pilots proliferate. Invoices arrive on time. And when the chief financial officer asks the one question that matters — what did we get back — the room goes quiet, or produces a slide of productivity anecdotes that would not survive an audit. MIT NANDA named this the adoption divide and attributed it not to model quality but to a learning and integration gap: tools that never adapt to the workflow, and organisations that never instrument the outcome.1
The instinct is to conclude that the technology has not delivered. That conclusion is wrong, and it is expensive, because it triggers exactly the two reactions that destroy value: panicked cost-cutting that kills the compounding pilots, or blind escalation of spend in order to catch up.
The evidence, read correctly
The claim that this technology does not pay is a misreading of the data. McKinsey's 2025 global survey of 1,993 leaders across 105 countries found that 88 per cent of organisations now use it and 74 per cent report first-year returns, yet only 39 per cent see enterprise-level EBIT impact and just 6 per cent qualify as high performers deriving 5 per cent or more of EBIT from it.20
The spread is the story. The average deployment returns about €3.40 for every €1 spent, while high performers return more than €9.48 — a roughly threefold gap driven not by better models but by reworking the work itself. High performers are 3.6 times more likely to make transformational change and far more likely to redesign workflows.20
Gartner expected at least 30 per cent of generative projects to be abandoned after proof of concept by the end of 2025, and warns that at least half overrun their budgets through poor architecture and missing operational know-how.21 Read together, this is a governance signal, not a technology verdict, set against a market IDC sizes at €581bn by 2028.22
The correct diagnosis is subtler than "it doesn't work" or "give it time." Cost and return are measured on different instruments. Cost is metered by the vendor, denominated in the vendor's units, and lands as a single legible number each month. Return is smeared across time, so this quarter's model rework saves next year's validation hours. It is smeared across functions, so a quality co-pilot reduces a deviation backlog that shows up in the operations numbers rather than the IT numbers. And it is smeared across counterfactuals you never ran. You are comparing a photograph to a weather system.
Why the paradox is a governance problem
If the return were simply absent, the rational move would be to stop. But the evidence from disciplined adopters is that the return is present and merely uninstrumented. The failure is that organisations deploy these capabilities the way they would install a coffee machine — buy it, plug it in and hope — rather than the way they would commission a piece of GMP equipment: define the intended use, qualify it, monitor it and tie it to a measurable output. The discipline Irish life sciences already applies to physical capital is exactly the discipline missing from its cognitive capital.
You would never install a bioreactor without an IQ, OQ and PQ protocol and a cost-of-goods model. Why would you install an intelligence pipeline without one?
That single analogy is the spine of this playbook. Everything that follows — the cost taxonomy, the financial model, the governance layers and the crossover analysis — is the qualification protocol and cost-of-goods model for cognitive capital. Make the cost governed and the return instrumented, and the paradox dissolves into a spreadsheet.
Three sentences to take to your board
- Our cost is fully knowable, so we will govern it to the cent.
- Our return is currently uninstrumented, so we will invest in measuring it before we invest further in spending.
- We will not scale any use case past a pilot until its unit economics and its value instrument both exist.
2. The anatomy of cost: a seven-band taxonomy
The vendor invoice shows one band of a seven-band cost structure. The other six are where budgets quietly die.
Ask most teams what their programme costs and they read the usage bill. That number is real, but it is typically 15 to 40 per cent of the true, fully loaded cost of a capability in production. To govern cost you must first see all of it.
| Band | What it is | Visibility | Typical share |
|---|---|---|---|
| 1 · Usage | Per-unit consumption charges, or the amortised cost of self-hosted inference | High, billed | 15–40% |
| 2 · Platform | Seats, orchestration tools, vector stores, observability, gateways | High | 5–15% |
| 3 · Infrastructure | Compute, storage, networking, development, validation and production environments | Medium | 10–20% |
| 4 · Data | Acquisition, labelling, cleaning, pipelines, retention | Low | 10–20% |
| 5 · Integration | Building, connecting to systems of record, ongoing maintenance | Low | 15–30% |
| 6 · Human review | Review, verification, operations labour, change management | Very low | 10–25% |
| 7 · Governance | Validation, compliance, audit, incident response, drift rework | Hidden | 5–20% |
Table 1. Indicative ranges for a production use case in a regulated environment. They do not sum to 100 per cent because mixes vary; the point is directional — bands four to seven routinely exceed band one.
An illustrative fully loaded decomposition for a regulated production use case puts usage at 28 per cent, platform at 10 per cent, infrastructure at 14 per cent, data at 14 per cent, integration at 19 per cent, human review at 9 per cent and governance at 6 per cent — meaning roughly 72 per cent of the true cost sits below the invoice waterline. A cost-control programme that only negotiates the usage bill is optimising a quarter of the problem. The leverage sits in data, integration and human review.
The two bands nobody budgets for
Band six, human in the loop. In a regulated setting most output is not shipped unread; it is reviewed. If a co-pilot drafts a deviation report in four minutes but a specialist spends twenty minutes verifying it, unit cost is dominated by the specialist's time, not by the tokens. The governance question is simple: does the tool reduce net human minutes, or merely relocate them?
Band seven, governance, risk and rework. Validation under GxP, EU AI Act conformity work, audit trails, incident response and the rework caused by model drift are real recurring costs. They are also costs a mature programme reduces over time by building reusable controls — which is precisely why they belong in the model, not in a footnote.
3. The financial model: eight equations that turn a hunch into a governance instrument
Eighteen equations stand between "we think this is worth it" and "we can prove it." Eight of them are load-bearing. Put them in a spreadsheet and you have a governance instrument.
Unit economics: the cost of one useful output
Everything starts with the cost per successful task. Not cost per token, not cost per call, but cost per output that actually cleared review and was used. Tokens are the vendor's unit; cost per successful task is yours.
| # | Equation | Formula |
|---|---|---|
| 1 | Cost per successful task | Ctask = (Cinfer + Creview + Cinfra) ÷ psuccess |
| 2 | Inference cost per attempt | Cinfer = (tin × πin) + (tout × πout) |
| 3 | Annual fully loaded TCO | TCOyr = Σb=1..7 Cb |
| 4 | Attributed benefit | Byr = Σj αj · vj · qj |
| 5 | Return on investment | ROI = (Byr − TCOyr) ÷ TCOyr |
| 6 | Simple payback | Tpayback = I0 ÷ (Bmo − Cmo) |
| 7 | Net present value | NPV = −I0 + Σ (Bt − Ct) ÷ (1 + r)t |
| 8 | Local-versus-hosted crossover volume | V* = Flocal ÷ (chosted − clocal) |
The success denominator in Equation 1 is the term teams forget. A model that is 30 per cent cheaper per call but succeeds 60 per cent of the time instead of 90 per cent is more expensive per useful output. Quality is a cost lever, not a separate axis. In Equation 2, output tokens are usually the dominant term — control verbosity and you control the bill.
For return, the instrument is benefit realised: the sum of labour hours redeployed, cycle-time value, error and rework avoided, and revenue enabled, each with an explicit and defensible attribution factor stating what share of the benefit is genuinely due to the capability rather than the process change around it. An ROI figure without its stated attribution factor is marketing, not measurement.
The governance test, in one line
A use case may scale past pilot if and only if cost per successful task is stable or falling, ROI is positive at a conservative attribution factor, NPV is positive at the corporate hurdle rate, and the value instrument exists. Miss any one and it stays a pilot.
4. The governance operating system: seven layers, four gates
Governance is not a policy PDF. It is a control structure with seven layers, each owning specific decisions, metrics and gates. Most governance failures are structural rather than moral: nobody owned the decision, so the decision was made by whoever held the corporate card.
| Layer | Owns | Key artefact | Cadence |
|---|---|---|---|
| L1 · Strategy | Why we invest, risk appetite, the return-against-cost thesis | Charter and investment thesis | Annual |
| L2 · Portfolio | Which use cases live or die, stage gates, prioritisation | Use case portfolio and scorecards | Quarterly |
| L3 · Financial | Budgets, unit economics, allocation, showback | Cost model and FinOps dashboard | Monthly |
| L4 · Risk | EU AI Act class, GxP validation, GDPR, audit | Risk register and conformity file | Continuous |
| L5 · Data | Provenance, quality, retention, access, residency | Data contracts and lineage | Continuous |
| L6 · Model and Ops | Model selection, hosting, evaluations, drift, versioning | Model registry and evaluation suite | Continuous |
| L7 · People | Skills, roles, review design, adoption, ethics | Capability plan and RACI | Ongoing |
Table 2. Strategy and risk appetite cascade down from L1; cost, evidence and risk signals flow back up from operations. Layer four, risk, is drawn widest because in a regulated Irish context it constrains every other layer — it is a rail on the whole track, not a gate at the end.
The single most important control: the stage gate
At layer two, every use case passes through the same four gates. No gate, no budget. This is how an unbounded field of pilots becomes a governed portfolio.
| Gate | Focus | Kill criterion |
|---|---|---|
| Gate 0 · Framing | Intended use defined, value hypothesis stated, EU AI Act risk class assigned | Kill if no owner |
| Gate 1 · Pilot | Thin slice built, cost per successful task measured, success rate measured | Kill if unit economics fail |
| Gate 2 · Validate | GxP and regulatory evidence assembled, controls in place, NPV positive | Kill if not defensible |
| Gate 3 · Scale | Cost allocation live, drift monitoring on, benefit instrument reporting | Reviewed quarterly |
5. The local-versus-hosted crossover: the core cost lever
The largest structural cost decision is make versus rent: pay per unit of consumption to a commercial service, or amortise your own hardware running open-weight models. There is a volume at which ownership wins, and it is knowable.
Commercial services are the right answer at low and spiky volume: near-zero fixed cost, frontier quality, no operations burden. But their marginal cost is constant — every unit costs the same as the last. Self-hosting inverts the shape: high fixed cost in hardware and operations labour, near-zero marginal cost. Somewhere the two lines cross. Below the crossover, rent. Above it, own. The crossover volume is Equation 8: V* = Flocal ÷ (chosted − clocal), where Flocal is the annual fixed cost of self-hosting (amortised hardware plus operations labour plus platform), chosted is marginal cost per unit on the commercial service, and clocal is marginal cost per unit self-hosted, mostly energy.
The numbers, as they stand in 2026
Renting has never been cheaper, or falling faster. Stanford HAI's 2025 index found the cost of inference for a mid-tier model fell from €18.40 per million tokens in November 2022 to €0.06 by October 2024, a reduction of roughly 280 times in about eighteen months, with task-dependent price declines of 9 to 900 times per year and hardware costs falling about 30 per cent annually.2 Current published list prices per million input and output tokens sit around €1.61 and €12.90 for flagship tiers, €4.60 and €23 at the top of the range, and €1.84 and €11 for mid-tier alternatives, per 2026 pricing trackers.10
Owning has never been more viable. Open-weight families have largely closed the real-world quality gap at roughly five to ten times lower per-unit cost. Self-hosting a 70 billion parameter class model on your own accelerators is estimated at five to twenty times cheaper than frontier list prices at high volume, with break-even around 5 to 10 million tokens per day.11
The three axes that move the crossover
| Axis | How it moves V* |
|---|---|
| Volume and steadiness | High, predictable volume favours ownership because the hardware stays busy. Spiky, low volume favours renting. |
| Privacy and residency | In pharma, sending intellectual property or patient-adjacent data to a third party can be a compliance cost or an outright bar. On-premise processing keeps data inside the boundary, which shifts V* effectively to zero for some workloads. |
| Quality headroom | If a mid-size open-weight model clears your success bar, ownership is viable. If only a frontier model passes, rent the frontier for those tasks and own the rest. The answer is usually a portfolio, not a monogamy. |
Hybrid is the mature default
Route cheap, high-volume, private tasks to owned open-weight models. Reserve the metered frontier service for low-volume, high-difficulty tasks where quality headroom justifies the premium. Governed routing between the two is the cost-control programme.
6. The implementation roadmap: four phases, four gates
A phased, gated path from first pilot to governed scale, designed so cost is controlled and value is instrumented at every step rather than retrofitted at the end.
| Phase | Focus | Gate | Timing |
|---|---|---|---|
| 0 · Foundations | Stand up layers one, three and four. Write the charter and risk appetite, build the empty cost model, assign use cases to EU AI Act risk classes. Establish the local sandbox so experimentation cost is near zero from day one. | Charter signed | Weeks 0–6 |
| 1 · Thin-slice pilots | Pick two or three high-volume, low regulatory-risk use cases. Build the thinnest slice that produces a real output. Instrument cost per successful task and success rate from day one. Kill anything whose unit economics do not converge. | Unit economics measured | Weeks 6–16 |
| 2 · Validate and harden | For survivors, assemble GxP and regulatory evidence, wire in human review controls, confirm NPV at the hurdle rate. Decide make-versus-rent per use case using the crossover. | Defensible, NPV positive | Weeks 16–28 |
| 3 · Govern at scale | Turn on cost allocation and showback, drift monitoring, and the benefit instrument. Move from project to portfolio: the quarterly layer-two review now runs the whole estate on evidence. | FinOps and drift live | Weeks 28 onward |
7. The cost maturity model: five levels
Most Irish organisations sit at level one or two and believe they are at level three. Maturity is not about how much you run — it is about how well you can answer questions about it.
| Level | State | Cost signal | What is missing |
|---|---|---|---|
| L0 · Shadow | Spending on personal cards, ungoverned | Invisible | Everything |
| L1 · Aware | Central bill exists, no unit economics | One total number | Attribution and unit cost |
| L2 · Measured | Cost tracked per use case | Ctask per use case | Return instrument |
| L3 · Governed | Stage gates, FinOps and ROI in place | ROI and NPV per case | Portfolio optimisation |
| L4 · Optimised | Portfolio managed, make-versus-rent routed | Marginal cost of value | Continuous frontier review |
The level-three self-deception
The most expensive place to sit is level two while believing you are at level three: dashboards full of usage metrics that look like governance but carry no return instrument and no kill gate. Usage is not value. If no use case has ever been killed at a gate for failing its economics, you are not yet governed. You are merely measured.
8. The FinOps metrics system: the dashboard that changes behaviour
If a number does not have an owner and a target, delete it. Vanity metrics measure activity, not stewardship. FinOps discipline has moved decisively into this territory: the FinOps Foundation reports that 98 per cent of organisations now manage this spend, up from 31 per cent in 2024, with getting to unit economics among the fastest-rising priorities.3
| Family | Metric | Owner | Target or signal |
|---|---|---|---|
| Unit | Cost per successful task by use case, month-on-month trend | Use case owner and FinOps | Flat or falling |
| Unit | Cost per 1,000 successful outputs | FinOps | Below the value per 1,000 |
| Unit | Review pass rate | Quality and use case owner | Meets the use case service level |
| Efficiency | Share of spend below the invoice waterline, bands four to seven | FinOps | Understood, not hidden |
| Efficiency | Idle hardware share for owned capacity, and cache reuse rate | Platform and operations | Idle below 20%, reuse rising |
| Value | Attributed annual benefit and ROI, with the attribution factor stated | Finance and use case owner | Positive at a conservative factor |
| Value | Payback period trend | Finance | Shortening |
| Governance | Showback coverage, share of spend allocated | FinOps | Toward 100% |
| Governance | Shadow spending incidents, and use cases past each gate | Layer-two portfolio owner | Incidents toward zero |
The perverse incentives to design out. A cost-per-call target invites teams to game verbosity down at the expense of quality, so watch the success rate alongside it. A raw ROI target invites optimistic inflation of the attribution factor, so require the assumption to be stated and reviewed. An idle-hardware target invites make-work to keep the cluster busy, so pair it with unit cost. Every single metric needs its counterweight on the same screen.
Start by showing each function its own cost. Behaviour changes the moment a team sees its number next to its peers. Only move to internal chargeback, where the cost actually lands on the function's budget, once the allocation model is trusted. A premature chargeback fight over imperfect allocation will stall the whole programme and drive spend back into the shadows.
9. The Irish and EU context: governance is a legal requirement, not an aspiration
In Ireland, governance is increasingly a legal requirement layered on an already heavily regulated industry.
The EU AI Act, live dates. In force since August 2024, with prohibited-practice bans applying from 2 February 2025 and general-purpose obligations from 2 August 2025. 2 August 2026 switches on penalty enforcement, Article 50 transparency duties and national market-surveillance sanctions.4 Under the 2026 Digital Omnibus, the heaviest Annex III obligations were deferred to 2 December 2027, and capabilities embedded in regulated products under Annex I to 2 August 2028 — breathing room, not a reprieve.5
Penalties are turnover-scaled. Up to €35m or 7 per cent of global turnover for prohibited practices, €15m or 3 per cent for high-risk non-compliance, and €7.5m or 1 per cent for supplying incorrect information.4 These are band-seven numbers that dwarf any usage bill.
Code of practice. If you build on general-purpose models, note the voluntary EU code of practice published 10 July 2025, with signatories listed from 1 August 2025, covering transparency, copyright and safety and security — the emerging baseline your vendors are expected to meet, and a due-diligence checklist for your own procurement.23
EMA, HPRA and GxP. The EMA's final reflection paper on use in the medicinal product lifecycle, adopted 9 September 2024 by CHMP and CVMP, sets a risk-based, human-centred bar from discovery through pharmacovigilance, with explicit weight on data quality and bias.7 Layered on GAMP 5 and EudraLex Annex 11 validation, the statement "the model changed and we did not revalidate" is an audit finding waiting to happen. Where the software is itself the device, the HPRA acts as national competent authority under EU MDR 2017/745 and IVDR 2017/746, and the IMDRF good machine learning practice principles set the lifecycle bar.24
GDPR and data residency. The EDPB's Opinion 28/2024, adopted 17 December 2024 at the request of the Irish Data Protection Commission, makes clear that model anonymity is assessed case by case and that unlawful training data can taint downstream use.6 For personal or IP-sensitive workloads this is a decisive argument for on-premise processing, since data stays inside the boundary, simplifying lawful basis and transfers.
Funding offsets the build. Ireland's research and development tax credit rose to 30 per cent for periods from 1 January 2024 and is legislated to 35 per cent for 2026, with the first-year payable threshold lifted to €75,000 — a combined benefit near 42.5 per cent with the trading deduction.9 Together with IDA Ireland and Enterprise Ireland supports, factor this offset into the upfront investment term in the payback equation.
Ireland is the second-largest exporter of pharmaceuticals in the European Union and the third-largest in the world, with about €99.9bn of pharmaceutical and medical exports in 2024, up roughly 29 per cent year on year and equal to about 45 per cent of all goods exports, across more than 90 companies employing some 50,000 people directly.8 The stakes of getting governance right here are national, not merely corporate.
The Irish adoption picture, by the numbers
The national data tells a story of rapid but deeply uneven adoption. Per the CSO's Information Society Statistics for enterprises in 2024, more than 15 per cent of all Irish enterprises adopted these tools in 2024, nearly double the 8 per cent of 2023. The average hides a chasm by firm size: 51.2 per cent of large enterprises, 25.1 per cent of medium and just 12.0 per cent of small.12 Among large firms, 30 per cent already automate workflows or assist decisions this way. Deloitte, with UCD Smurfit, found 62 per cent of leading Irish companies adopting, rising to 94 per cent in technology, media and telecommunications, that one in four Irish firms now has a chief officer for the domain, and ranked Ireland first in EMEA for strategy integration.13 Against this sits the national ambition of the refreshed strategy: 75 per cent of enterprises using these technologies, cloud and big data by 2030.14
The talent constraint is the binding one
In Ireland the scarcest input is not compute, it is people. The Expert Group on Future Skills Needs, with NIBRT, projects 14,000 to 26,000 additional biopharma jobs over five years on a base of 50,000, while warning that graduate inflow could fall short by about 3,000 roles every year, with acute shortfalls in digital and data scientists and in chemical and digital process engineers.18 This has a direct consequence for the cost model: it inflates band six, human review, and band five, integration and engineering, wage rates. That is precisely why a governance programme which reduces net human minutes per output, rather than merely cutting the usage bill, is the higher-leverage play in this market.
The Irish funding stack
| Support | Body | What it offsets |
|---|---|---|
| R&D tax credit, 30% rising to 35% for 2026 | Revenue | Qualifying build work, combined benefit near 42.5% |
| Discovery voucher and digital discovery grant | Enterprise Ireland | Structured feasibility and assessment at gate zero |
| Grow Digital voucher | Local Enterprise Office | Small-business digital tool adoption |
| European Digital Innovation Hubs, phase two, 2026–2029, €23m | CeADAR, FactoryXChange, Data2Sustain, ENTIRE | Test-before-invest, skills and technical expertise |
| Industry 4.0 and readiness assessment | Irish Manufacturing Research | Manufacturing readiness and deployment, more than €6m in live projects |
Table 5. Few jurisdictions match Ireland's density of state supports for exactly the band-three and band-five build costs this model tracks. Fold these into the upfront investment before concluding a use case fails its payback test.
IDA Ireland reported €13.2bn of client capital investment and a record €1.9bn in research, development and innovation in 2024, with foreign direct investment supporting 544,619 jobs. The biopharmachem sector alone reports about €116bn in exports and 80,000 direct and indirect jobs.16,19 Since 1 August 2024 national research funding runs through Taighde Éireann, Research Ireland, backing centres including Insight, ADAPT and CeADAR, which alone secured €5.675m under the second phase of the European Digital Innovation Hubs.25 For a life sciences firm these are not academic curiosities. They are subsidised routes to de-risk the band-four and band-five build before it reaches your own profit and loss account.
10. Risk register: the failure modes of ungoverned spend
Ungoverned spend fails in predictable ways. A governed register, scored, owned and reviewed on a cadence, turns vague anxiety into specific controls. The point is not the register's existence but its liveness — a risk with no owner and no last-reviewed date is decoration.
| Risk | Failure mode | L×I | Control | Owner |
|---|---|---|---|---|
| Shadow spending | Ungoverned spend and data leakage | H×H | Central gateway, showback, gate zero for every use case | L2 portfolio |
| Runaway usage | Verbose outputs, loops, no cap | M×H | Spend caps per use case, output length limits, alerts | FinOps |
| Quality collapse | A low success rate inflates cost per task | M×H | Evaluation suite, success-rate service level, model routing | Quality |
| Model drift | Silent degradation leading to rework | M×M | Drift monitoring, versioned registry, scheduled re-evaluation | Operations |
| Vendor lock-in | Price rises with no exit | M×M | Abstraction layer, and an owned open-weight fallback | Architecture |
| Compliance gap | Unvalidated model in a GxP or high-risk path | L×H | Layer-four rail, conformity file, revalidation on change | Regulatory |
| Data leakage | IP or personal data sent to a third-party service | M×H | Loss prevention at the gateway, local models for sensitive data, residency rules | DPO |
| Over-automation | Human oversight removed where it mattered | L×H | Review by design on consequential decisions, statutory oversight duties | Function lead |
| Concentration | One model or vendor underpins many use cases | M×M | Diversified routing, documented fallbacks, exit test | Architecture |
Cadence. Runtime risks — runaway consumption, drift and quality — are monitored continuously and reviewed monthly at layer three. Structural risks — lock-in, concentration and compliance — are reviewed quarterly at layer two. The full register is re-scored at the annual layer-one strategy review. The three that most often kill programmes quietly are shadow spending, quality collapse and the compliance gap. Govern those first.
11. Three case vignettes from Irish life sciences
Three composite Irish scenarios, worked end to end through the equations above. All figures are illustrative: plausible, internally consistent, and clearly not any real firm's private data.
A · The Cork contract manufacturer: document extraction at volume
A contract manufacturer extracts structured data from batch records and certificates of analysis: 40,000 documents a year, each roughly 8,000 input and 1,500 output tokens. The incumbent design uses a frontier hosted service plus heavy specialist review. The question is whether a fine-tuned open-weight model on premise beats it.
| Component, per document | Before: hosted, heavy review | After: on premise, light review |
|---|---|---|
| Inference cost | €0.05 | €0.01 |
| Human review | 12 min × €60/h = €12.00 | 5 min × €60/h = €5.00 |
| Allocated infrastructure | €0.20 | €0.55 |
| Success rate | 0.90 | 0.94 |
| Cost per successful task | €13.61 | €5.91 |
Net cost per successful task falls by about 57 per cent, and note where the saving lives — almost all of it is band six, review minutes, not the usage bill. On pure inference the crossover barely matters at this volume, since roughly 570 million tokens a year sits below a cost-only threshold. What tips the decision to own is GDPR and intellectual-property residency, plus the review-time gain from a model tuned to the document format. The hosted frontier service is retained only for the 2 per cent of genuinely ambiguous documents — the hybrid routing described in Part 5.
B · The hospital group: clinical administrative drafting
Two hundred clinicians save 30 minutes a day on administrative drafting, 220 days a year: 22,000 hours a year, valued at €70 an hour loaded, with a deliberately conservative attribution factor of 0.4. That produces an attributed benefit of €616k. Annual TCO is €180k — fully loaded on-premise cost: hardware amortisation, operations, validation and support, with data sensitivity making local processing mandatory under GDPR. With an upfront build of about €150k and net monthly benefit near €36k, the return is 242 per cent and payback is 4 months, with NPV comfortably positive at any sane hurdle rate.
The lesson: even on a conservative attribution and a local-only constraint, the case clears on labour redeployment alone, provided the benefit is instrumented through time-and-motion measurement rather than assumed.
C · The mid-cap biologics manufacturer: a quality co-pilot
A quality co-pilot drafts deviation and corrective-action documentation. Volume is medium and regulatory load is high, so band seven dominates. The instructive finding is that the cheaper model is the wrong lever.
| Lever | Action | Annual effect on cost per task |
|---|---|---|
| Cut the model price | Swap the frontier hosted service for a cheaper model | −€0.04, negligible |
| Redesign the review | Structured templates and confidence flags cut review from 20 to 12 minutes | −€8.00, dominant |
In a regulated setting, cost control lives in bands six and seven, not band one. The winning move is not procurement but process: a review workflow the tool is designed to accelerate, and a validation approach that makes drift cheap to catch. Chase the usage bill here and you optimise a rounding error while the real cost — specialist minutes and rework — runs untouched.
12. Templates: the board memo, the equation card and the RACI
Adopt tomorrow's artefacts, written out in full. Copy them into your own stack and you have the skeleton of a governed programme.
The one-page monthly board memo
- Thesis. Cost is knowable and governed to the cent; return is being instrumented before further spend.
- Portfolio. Use cases live, split across gate one, gate two and gate three; number killed this quarter and the reason.
- Unit economics. Blended cost per successful task this month against last; best and worst use case; success-rate trend.
- Value. Attributed benefit year to date with the attribution factor stated; portfolio ROI; use cases below the NPV threshold.
- Make versus rent. Decisions taken this month and the crossover volume that drove them.
- Risk. Top three register movements and any regulatory items.
- Asks. Decisions and budget needed from the board.
Cost model spreadsheet specification
Eight tabs. Inputs: holding prices, wage rates, hurdle rate and grant offset. Bands: the seven-band build-up from Table 1. Unit: Equations 1 and 2. TCO: Equation 3. Value: Equations 4 and 5 with the attribution factor as an explicit, reviewable input. Capital: Equations 6 and 7. Crossover: Equation 8. Dashboard: the Part 8 metric set. Every headline number on the dashboard traces to a formula, and every formula traces to an input. No hard-coded results.
Gate-zero use case intake form
A use case may not consume budget until this one-pager is complete: a named accountable owner; the intended use and value hypothesis, meaning what output and what it is worth; the value instrument, meaning how the benefit will be measured; the EU AI Act risk class; the data class and residency, meaning whether personal or IP data is involved and whether local processing is required; the make-versus-rent hypothesis; and the success and kill criteria for the pilot. No owner, no intended use and no instrument means no gate zero.
Governance charter and RACI
| Layer | Accountable | Responsible | Consulted |
|---|---|---|---|
| L1 Strategy | Board and chief executive | Chief digital officer | Finance, quality |
| L2 Portfolio | Chief digital officer | Portfolio lead | Function leads |
| L3 FinOps | Chief financial officer | FinOps lead | IT, function leads |
| L4 Risk and compliance | Chief quality and regulatory officer | Regulatory affairs, data protection officer | Legal, IT security |
| L5 Data | Chief data officer | Data engineering | Data protection officer, quality |
| L6 Model and operations | Head of machine learning | Operations | Security, quality |
| L7 People and change | Chief human resources officer | Learning and development, change leads | Function leads, unions |
A closing word
The organisations that will thrive in Irish life sciences are not those that spend the most, nor those that wait for certainty. They are the ones that treat cognitive capital with the same discipline they already bring to physical capital, governing a knowable cost while patiently instrumenting an uncertain return. That is not a technology programme. It is a management one.
Selected sources
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, reported August 2025
- Stanford HAI, 2025 AI Index Report, chapter 1, Stanford University
- FinOps Foundation, State of FinOps 2025 and 2026
- European Commission, EU Artificial Intelligence Act implementation timeline and penalties, 2026
- European Commission, EU Digital Omnibus, postponed high-risk deadlines, 2026
- EDPB, Opinion 28/2024 on data protection aspects of AI models, December 2024
- EMA, reflection paper on the use of artificial intelligence in the medicinal product lifecycle, September 2024
- IDA Ireland, IPHA and BioPharmaChem Ireland, Irish pharmaceutical sector data, 2025
- Revenue and KPMG, Ireland research and development tax credit, 2026
- CSO, Information Society Statistics, Enterprises 2024, released February 2025
- Deloitte Ireland and UCD Michael Smurfit, Irish adoption research, 2025
- McKinsey & Company, The State of AI, global survey 2025
- Gartner, generative project abandonment and budget overrun research, 2025
- IDC, worldwide market forecast, 2025
- EGFSN and NIBRT, Skills for Biopharma 2024, April 2024
Adapted from Priya Life Sciences' 2026 intelligence brief, The Cost-Governed Enterprise: A Digital Transformation Playbook for Implementing Advanced Computation in Irish Life Sciences, and for Financially Governing Its Cost Before It Governs You, by Sreepriya Prasannan, Founder and Editor, Priya Life Sciences.
