Ai implementation is reshaping how we work. Every AI vendor’s website claims their product produces enormous ROI. Every AI consultancy’s case studies show eye-watering returns. In particular, every internal AI proposal is built around a financial case that, on closer inspection, contains assumptions that wouldn’t survive a finance director’s first review.
The honest position is more nuanced. AI deployments can produce excellent returns – but only when the calculation is done properly, with realistic inputs and including the costs that don’t appear on the vendor’s quote. Building this calculation properly is the single best way to separate AI initiatives that will succeed from those that will quietly underperform.
The four AI implementation cost categories that matter
A complete AI ROI model has four cost categories. Most internal proposals account for the first one and ignore the rest.
1. Direct deployment cost. The model API spend, the development work to integrate it, the infrastructure to run it. This is what gets quoted. It is typically 30-50% of the true total.
2. Integration and data plumbing. The work to get data into a form the AI can use and the work to get its output into the systems that need it. For most enterprise deployments, this is the largest single cost – often exceeding the AI integration itself by 2-3x.
3. Evaluation and quality infrastructure. Building the test sets, monitoring, human review pipelines and feedback loops that keep the system working. Frequently underestimated, sometimes ignored entirely. The deployments that ignore this layer are the deployments that quietly fail in production six months in.
4. Change management and training. Getting the people who will use the AI to actually use it, change their workflows, trust its outputs appropriately. For internal-facing AI, this is often the deciding factor between a successful deployment and an unused one.
A complete cost model accounts for all four. A typical proposal that costs £40k for the AI build will, in reality, cost £80-120k once integration, evaluation and change are properly funded. This is not bad news – the returns are usually still very strong – but it has to be accounted for honestly to avoid surprises.
The three benefit categories
equally, the benefits of AI deployment fall into three categories, only one of which is usually quantified.
1. Direct labour displacement. The most commonly modelled. If a process took 10 hours a week and the AI reduces it to 2 hours, that’s 8 hours a week reclaimed. Multiply by hourly cost, annualise, compare to investment.
This is real, but it is also the most easily over-claimed. People rarely fully repurpose the time saved – some of it is absorbed by other work, some by recovery from increased throughput elsewhere, some lost. Realistic capture rate is usually 60-80% of the theoretical saving.
2. Throughput expansion. Often more valuable than labour saving. A team that could process 100 requests per week can now process 300, without expanding headcount. This unlocks revenue or capacity that didn’t exist before. For sales, support, content production and operational processing, this is typically the largest source of value.
3. Quality improvement. Hardest to quantify, often most valuable. Better-quality outputs, faster turnaround, more consistent customer experience. The downstream effects – retention, referral, conversion – are real but diffuse. We try to capture them with proxy metrics rather than guessing at attribution.
A serious ROI model looks at all three. The single-category model – usually labour displacement only – almost always understates the true return.
The time-to-value question
The other under-modelled variable is time. AI deployments rarely produce their full benefit on day one. There is a ramp:
- Months 1-3: Pilot phase. Limited deployment, heavy human oversight, lots of iteration. Net cost often exceeds net benefit.
- Months 4-6: Production rollout. Benefits start to accrue but at a fraction of steady state. Costs of monitoring and refinement are still elevated.
- Months 7-12: Steady state. Full benefits realised, costs stabilised. This is the period the ROI model should be anchored on.
- Year 2+: Compounding. The system gets better, the team’s confidence grows, additional use cases get layered on. ROI accelerates.
A 12-month ROI calculation tends to underestimate; an 18-24-month calculation gets closer to reality. For deployments designed to compound, three years is often the right horizon.
The framework we use with clients
when we scope AI work, we structure the financial case around five questions:
- What is the current cost of the process or capability we’re addressing? Time, money, opportunity cost, error rates, customer impact. Quantified as honestly as possible.
- What is the realistic best-case post-deployment state? Not the marketing number – the genuine, achievable performance after 12 months of operation, including the parts that won’t automate.
- What is the full investment required? All four cost categories, time-phased, with a contingency for the unknown work that always emerges.
- What does steady-state ROI look like? Once the deployment is mature, what is the ongoing benefit minus ongoing cost? This is the number that should drive the decision.
- What is the breakeven point and the risk profile? When does the deployment pay for itself? What could delay or prevent that? What’s the contingency plan?
A deployment that survives this analysis is one worth pursuing. A deployment that doesn’t is one that would have failed expensively if pursued anyway.
Where AI returns are reliably strong
Across the deployments we’ve helped clients model, the patterns of strong return are consistent:
Document-heavy operational processes. Insurance, legal, procurement, accounts payable, compliance review. AI’s ability to read, classify, extract and summarise unstructured documents at speed transforms these workflows.
Customer support tier 1 and 2. Well-deployed AI handles 50-80% of common queries fully and accelerates resolution on the rest by surfacing the right information to human agents.
Content production at scale. Marketing teams producing 10x the volume with the same headcount, with quality maintained through structured workflows.
Sales intelligence and research. The hours sales teams spend researching prospects, preparing for meetings, drafting proposals – AI eats into this dramatically without affecting quality.
Internal knowledge access. AI assistants over internal documentation, policies and historical data save hours of search and asking-around per employee per week.
Where returns are consistently weaker: highly creative work where the differentiation matters more than the output volume; tasks already heavily automated by deterministic tools; processes where the human judgement was the value, not the throughput.
The strategic point
AI implementations have a real, often substantial, financial return. Capturing it requires honest cost modelling, realistic benefit projection and the discipline to commit to the unglamorous parts of the deployment alongside the exciting ones.
the businesses winning with AI are not the ones with the most ambitious projections. They are the ones whose projections turned out to be roughly right.
Modelling the case for an AI initiative and want a sanity check on the numbers? We help clients build defensible business cases that finance directors actually believe. Get in touch.