What an AI roadmap actually looks like for a mid-size company

Why most AI strategies stall

Most mid-size companies we speak with already have an AI strategy of some kind. Often it is a presentation from a workshop: a list of use cases, a maturity model, perhaps a vendor shortlist. Six months later, very little has changed on the ground.

The problem is rarely ambition. It is that the strategy was written at the level of technology (“we should use generative AI for customer service”) instead of the level of work (“our team spends three days a month matching supplier invoices to deliveries”). Technology-first plans are hard to price, hard to prioritise and easy to postpone.

A roadmap that gets executed looks different. It is specific, it is ranked, and it is tied to numbers someone in the business already cares about.

Step 1: Map the work, not the technology

We start with a short discovery across the teams that move money, inventory, documents or decisions: finance, operations, procurement, sales and customer service. In each one, we look for four patterns:

Re-keying. The same information typed into more than one system, or copied from PDFs, emails and handwritten notes.
Waiting. Work that stops until someone finds a document, gets an approval or reconciles two numbers.
Searching. People hunting through folders, contracts or old emails to answer routine questions.
Judgment without data. Decisions made on instinct because the data exists, but not in one place.

Each pattern becomes a candidate. A week of conversations in a typical mid-size company surfaces 15 to 30 of them. The goal is not to find “AI use cases”. It is to find expensive friction. Some of it will be solved by AI, some by better software, and some by a process change that costs nothing.

Step 2: Score every opportunity

Next, every candidate is scored on three factors. We keep the scoring simple enough for the leadership team to debate it in a single meeting.

Factor	What we ask	Score
Impact	Hours saved, errors avoided, revenue protected or faster decisions, expressed in money	1 to 5
Effort	Integration work, change management, and the number of teams involved	1 to 5, lower is better
Data readiness	Does the data exist, can we access it, and is it reasonably consistent?	1 to 5

Plot impact against effort, and size each opportunity by data readiness. The first project comes from the top left: high impact, low effort, data ready.

Data readiness is the factor most often skipped, and the one that most often sinks projects. A brilliant idea that depends on three years of data nobody recorded is a second-phase project, not a first one.

Step 3: Check the data foundation

For the top five or so opportunities, we look underneath. Not a full data audit, just enough to know whether a project can be built in weeks rather than months.

Where does the data live today: an ERP, accounting software, spreadsheets, email or paper?
Who owns it, and who can grant access?
How much history exists, and how consistent is it?
Are there documents, such as invoices, contracts or reports, that need to be read rather than queried?
What privacy, data residency or regulatory constraints apply?

This step often changes the order. A project that looked second best can jump to first because its data is already clean and sits in one system.

Step 4: Sequence into three horizons

We then lay the work out in three horizons. They are about dependency rather than calendar ambition: each one builds the foundation the next one needs.

0 to 90 daysProve value
One or two systems in production for a real team: typically document processing, a decision dashboard or a workflow automation.
3 to 9 monthsConnect
Link the first systems to core data sources, add the next priorities, and put monitoring and ownership in place.
9 to 18 monthsScale
Copilots and agents that work across departments, built on data and integrations that now exist.

Each horizon earns the next. Nothing moves into a later horizon until the earlier one is live and used.

The discipline here is saying no. Everything that does not fit in the first horizon waits, and is scored again once the first systems are live. Real usage almost always reshuffles the list.

Step 5: Build a business case finance will sign

Each project in the first horizon gets a one-page business case. It should be boring and verifiable:

Baseline. How the work is done today, how long it takes, and how often errors happen.
Target. What changes, and how it will be measured.
Cost. Build, run and support, including the time your own people will spend.
Payback. When the savings or gains cover the cost.

Illustrative example

A distribution company processes 1,200 supplier invoices a month. Each takes about six minutes to key in and check: around 120 staff hours a month. An extraction system with human review brings that to roughly a minute and a half per invoice, and flags price mismatches before payment. That frees about 90 hours a month, before counting the mismatches it catches. These figures are illustrative. In a real roadmap they come from your own baseline.

If a project cannot be described this way, it is not ready for the first horizon.

Governance that does not slow you down

A mid-size company does not need an AI ethics committee before its first project. It does need a few rules, agreed early:

A person reviews any AI output that moves money, changes records or reaches a customer, until accuracy is proven.
Every AI answer shows its source, so people can check it in seconds.
Access follows existing roles. The system never shows anyone data they could not already see.
Usage and accuracy are logged, so you can prove the system works and notice when it drifts.

What the roadmap document should contain

The roadmap itself is short. Ours usually fits in ten pages:

The ranked opportunity list, with scores and reasoning.
A data readiness summary for the top priorities.
The three-horizon plan, including what is deliberately not being done yet.
One-page business cases for the first horizon.
Governance rules, and who owns what.
A review date, usually 90 days out, when the plan is scored again against reality.

Where to start

If you do one thing this month, list the ten most repetitive, document-heavy or approval-heavy tasks in your company, and estimate the hours each one consumes. That list is the raw material for a roadmap, and it costs nothing to make.

If you would like a second pair of eyes on it, that is exactly the conversation we have in a first consultation.