How to prioritise AI integration in your company: a quantitative model

Do you know how you are going to implement AI in your company? Not which tool. How. And above all: where first?

The honest answer from most people is a gesture. "Where it hurts most." "Customer service, it is overwhelmed." "Whatever the competition is doing." That is not a decision. It is a hunch in a suit.

Deciding where AI goes in a company is a quantifiable problem. It has variables, it has formulas and it has a rankable result. This article is that model. With one warning up front: a model is not a crystal ball. It orders your decisions, it does not guess them. More than giving you exact numbers, it gives you a disciplined way of estimating them.

You will not find here which language model to choose. You will find how to know, with numbers, which process to automate first and how much it is worth doing.

I have been working with these tools since before ChatGPT launched commercially. And if I have learned anything, it is that the problem is almost never the AI. The problem is that people decide without measuring. So let us measure.

Phase 0 — The unit of analysis: the process, broken down

Before any number, you have to look inside each process. Every process, however complex it looks, breaks down into four components:

• Data input — what information comes in and in what form.
• Connection nodes — which systems are involved and how they communicate.
• Processing — what transformation or decision happens.
• Expected output — what result is produced and how it is validated.

This is not decorative. It is the basis of everything that follows, because the feasibility of automating a process depends on the quality of these four components. A process with dirty data at the input, or with no connection between systems, is not an AI problem. It is a process problem. And those get solved first, without AI.

Hence the standing principle: First optimizer, second AI.

Phase 1 — Inventory and classification

List every process in the company. Then classify them by nature:

• Repetitive: clear rules, high frequency.
• Digital: the information already lives in software.
• Physical but digitisable: manual today, but convertible.

Only repetitive processes that are already digital, or that you can digitise, move to the next phase. The rest fall outside the analysis. Not because they do not matter, but because they are not candidates today.

Phase 2 — The cost of manual work (CM)

Here comes the first hard number. The annual cost of running a process by hand:

CM = (V × t × Ch) + (V × e × (tr × Ch + Cext))

Where:

• V = executions per year
• t = hours per execution
• Ch = loaded cost per hour (salary + social charges + overhead)
• e = error rate (0 to 1)
• tr = rework hours per error
• Cext = external cost per error (refunds, penalties, customer lost)

The first term is the direct cost of the work. The second is the hidden cost of the errors. The one almost nobody adds up, and the one that often weighs most.

Phase 3 — The Feasibility Factor (FV)

The cost tells you how much it hurts. It does not tell you whether it can be solved. Here the four components come back. Score each one from 0 to 1 according to how ready it is for automation:

• Input: is the data digitised and structured? (0 = paper/chaos, 1 = clean and structured)
• Connection: do the systems communicate? are there APIs or integrations? (0 = islands, 1 = connected)
• Processing: is the logic deterministic and documented, or does it need human judgement that is hard to encode? (0 = pure judgement, 1 = clear rules)
• Output: is the expected result clear and verifiable? (0 = ambiguous, 1 = defined and measurable)

The Feasibility Factor combines all four. And here an important decision: do not use a simple average. The four components are a chain. If one fails, the process fails. That is why a geometric mean is the right choice:

FV = (Ce × Cx × Cp × Cs) ^ (1/4)

Where Ce, Cx, Cp, Cs are the scores for input, connection, processing and output.

Why geometric? Because it penalises imbalance. A process with three perfect components and one at zero gives FV = 0. With no data at the input, no AI is worth anything. The weakest link rules.

One important clarification: the FV measures technical feasibility, not whether it is safe to run. A process can be highly automatable and still touch sensitive decisions — money, case files, regulation. That does not lower its FV, but it does require a governance layer: human oversight, traceability and explicit limits. We come back to this in the case.

Phase 4 — The three impact vectors

The classic mistake is prioritising by cost. Automating the most expensive thing. But the most expensive is not always the most valuable. The value of optimising a process is spread across three vectors, and you have to add them up.

1. Resource saving (AN)

How much of that manual cost you recover, minus what it costs to run the solution:

AN = (α × CM) − Co

• α = automatable fraction of the process (0 to 1)
• Co = annual operating cost of the solution (APIs, licences, maintenance, human oversight)

Watch out for α: it is almost never 1. Automating 100% of a process is rare. The normal case is that the AI takes the bulk and an oversight layer remains.

And watch out for something more important: this saving is only cash if it reduces a real cost or avoids a hire. If the team is still on payroll doing something else, you did not save money: you freed capacity. They are different things.

• Cash saving = a real cost reduction or a hire avoided.
• Freed capacity = hours that can be reassigned to higher-value work.

Freed capacity is worth something, but it only turns into money if you reinvest it in something that reduces cost or generates revenue.

2. Impact on revenue (ΔR)

The additional annual revenue attributable to optimising the process. You estimate it, you do not invent it. And you discount it:

ΔR = (measurable improvement) × (unit value) × attribution factor

The attribution factor (0 to 1) recognises that not all of the improvement is down to the AI: there is seasonality, marketing, pricing, a thousand things. You rarely attribute 100%.

3. Impact on customer service (ΔCX)

The hardest to monetise, but not intangible. You approximate it through retention:

ΔCX ≈ Δchurn × Customers × LTV × confidence factor

• Δchurn = expected reduction in the churn rate
• LTV = customer lifetime value, in net contribution, not gross revenue
• confidence factor (0 to 1) = how much you believe the estimate

A general rule for the two soft vectors: if you do not observe it directly, it is a hypothesis. Estimate it low, document it and validate it after implementing. A model that treats its assumptions as facts is not rigour: it is marketing.

Total annual value (with ΔR and ΔCX already adjusted):

VT = AN + ΔR + ΔCX

Phase 5 — The Integration Priority Index (IPI)

You now have, for every process: how much it costs, how feasible it is and how much it is worth. Now you rank them. If you could monetise the three vectors with confidence, the route is direct:

IPI = VT × FV / Ci

Ci = implementation cost (initial investment, one-off).

And here a clarification that saves your argument from being knocked down: the IPI is not an ROI. It mixes savings, estimated revenue and customer experience, adjusted by feasibility. It is a prioritisation index — closer to an annual benefit-cost ratio — for ranking opportunities before deciding where to invest analysis, design and implementation. The financial ROI comes afterwards, once you have chosen where to look.

When not everything can be monetised — typical with customer service — use the multi-criteria version. Normalise each impact vector to a 0–1 scale and weight it according to the business strategy:

IPI = (wₐ·Îₐ + w_f·Î_f + w_c·Î_c) × FV / Ê

• Îₐ, Î_f, Î_c = normalised impacts (0 to 1) on savings, revenue and service
• wₐ, w_f, w_c = strategic weights that sum to 1
• Ê = normalised implementation effort

The weights are not technical. They are strategic. A company in efficiency mode raises wₐ. One in growth mode raises w_f. That is where your quantitative analysis meets the direction of the business.

Two validations before executing:

Payback (months) = Ci / (AN / 12)
ROI over n years = (n × AN − Ci) / Ci

With an honest caveat: that payback uses the equivalent operating saving. It is only a cash payback if that capacity turns into reduced cost, an avoided hire or revenue. And priority is not the same as immediate profitability: if the IPI is high but the payback is four years, look again.

A case: a university that wants to start

A mid-sized university decides to integrate AI. On the table, two candidate processes. Instinct says "let us start with the easy one". The model says otherwise. Illustrative figures, annual.

Process A — Handling admissions and student enquiries

Broken down into its four components:

• Input: enquiries by email, portal and chat about requirements, deadlines, grants and enrolment.
• Connection: admissions CRM + academic management system + regulations.
• Processing: classify and answer; deterministic FAQs + RAG over the regulations; escalation to a person in complex cases.
• Output: a verifiable answer, with a human handoff where it applies.

V = 40,000 · t = 0.2 h · Ch = 18 · e = 0.04 · tr = 0.5 h
CM = (40,000 × 0.2 × 18) + (40,000 × 0.04 × 0.5 × 18) = 144,000 + 14,400 = 158,400
α = 0.65 · Co = 25,000 · Ci = 60,000 → AN = (0.65 × 158,400) − 25,000 = 77,960
FV = (0.8 × 0.7 × 0.7 × 0.9) ^ (1/4) ≈ 0.77

That AN is an equivalent operating saving: capacity freed up in the team. It becomes cash if it cuts overtime, avoids hires at enrolment peaks or gets reinvested in recruitment.

And here is what a cost-only analysis would never see. First, the raw estimate:

• ΔR (revenue): better answers and 24/7 availability → ~40 more enrolments a year × 3,000 in net contribution = 120,000 gross.
• ΔCX (retention): −0.3 pp of dropout across 5,000 students = 15 students × 9,000 in net LTV = 135,000 gross.

They do not double-count: ΔR is new recruitment; ΔCX is retention of students already enrolled. Different populations.

Now the adjustment, because these are hypotheses, not facts:

ΔR adjusted = 120,000 × 0.6 (attribution) = 72,000
ΔCX adjusted = 135,000 × 0.5 (confidence) = 67,500
VT = 77,960 + 72,000 + 67,500 = 217,460
IPI = (217,460 × 0.77) / 60,000 ≈ 2.79
Payback (on the operating saving) ≈ 9 months

With nothing discounted, A’s IPI would be 4.27. With a prudent discount, 2.79. We keep the credible number.

Process B — Issuing certificates and document validation

V = 12,000 · t = 0.15 h · Ch = 16 · e = 0.03 · tr = 0.4 h
CM = (12,000 × 0.15 × 16) + (12,000 × 0.03 × 0.4 × 16) = 28,800 + 2,304 = 31,104
α = 0.9 · Co = 6,000 · Ci = 25,000 → AN = (0.9 × 31,104) − 6,000 ≈ 21,994
FV = (0.9 × 0.8 × 0.95 × 0.95) ^ (1/4) ≈ 0.90
ΔR ≈ 0 (administrative process) · ΔCX = 5,000 × 0.5 ≈ 2,500
VT ≈ 24,494
IPI = (24,494 × 0.90) / 25,000 ≈ 0.88

The decision

Process B is cleaner, more deterministic and almost 100% automatable. Its feasibility is higher (0.90 against 0.77). It is the technical slam dunk. And even so the model says: start with A.

An IPI of 2.79 against 0.88. More than three times higher priority, even after discounting the soft impacts.

Because ease does not generate value. Impact does. And in a university the value is not in issuing certificates faster: it is in turning applicants into enrolments and in those enrolments not falling through.

Here is the heart of it: technical ease should not decide the priority. At most, it should decide the learning sequence. That a process is easy to automate is an engineering argument, not a strategic one. And an organisation should not let the easy option — or the technical department — define where the business is played out.

Priority is not the same as first pilot

The IPI tells you where the biggest economic opportunity is. Not necessarily where to execute first. Sometimes it is worth starting with the simple process — B — to test the integrations, produce a quick win and lower internal resistance, and reserve the serious investment for the higher-impact one — A.

• Strategic priority: A.
• Low-risk pilot (optional): B.
• Main investment: A.

And one brake that is not optional. Process A touches admissions, grants, enrolment and regulation. An error there is not a badly answered ticket: it is incorrect information about requirements, an administrative or legal problem, a complaint, reputational damage. So: no process with legal, academic or reputational impact runs without a human handoff, traceability and explicit limits on what it can answer. Technical feasibility is not enough.

A final warning about the numbers: ΔR and ΔCX are the most sensitive to your assumptions — the attribution and confidence factors are the lever. That is why you estimate them low and validate them against reality afterwards. Even so, you would have to be very pessimistic for A to stop winning: halve them again and its IPI is still above B’s.

From the numbers to the strategy

The model does not automate anything. It ranks. It tells you where to start, how much it is worth and how feasible it is. From there, the strategy is simple to state and disciplined to execute: you choose where the biggest opportunity is (the highest IPI), you decide the execution sequence according to the risk you can take, you measure the real result against the estimate, you adjust your assumptions, and you move to the next one.

And only then — at the end of the analysis, not the beginning — do you choose the tool. Because by that point you already know exactly which problem it solves, how much solving it is worth and how ready the ground is.

Almost everyone believes that implementing AI is a technology decision. It is a decision about analysis, business and risk. First you measure your processes. You break them down. You calculate their cost, their feasibility and their value. AI is what you apply afterwards, to the process the model pointed at.

Because value does not appear when you plug in a tool. It appears when you know exactly which process deserves to be transformed, how much transforming it is worth and what risk you are willing to take.

That is why I will repeat it, now with more reason:

First optimizer, second AI.

And "optimizer", here, means one concrete thing: the person who runs the numbers — and takes the risk with their eyes open — before switching on the machine.

If you want to apply this model to your operation — measure your processes, calculate their IPI and decide with numbers where AI starts — that is exactly where we begin in our AI consulting for companies and in every process automation project. Let us talk about your operation and we will tell you, with numbers, which process deserves to go first.