What an AI transformation roadmap looks like for a business under 500 people

An AI transformation roadmap for a business under 500 people is six phases in order: baseline, guardrails, data, conversation automation, workflow redesign, agentic operations. Week one is measurement, not software. Data usually takes longest. In McKinsey’s 2025 survey, only 29% of respondents from companies under $100 million revenue had reached the scaling phase.

Sources last checked: 23 September 2026. Every external figure on this page links to the publisher that produced it, and was re-read at that source before publication.

• The sequence: the Baseline-First Roadmap. Six phases, each with a first action and a finish state you can check.
• Week one: write down the current value of one metric per process you might automate, and list the AI tools staff already use.
• Longest phase: phase 3, getting the data for your first use case fit to use.
• Where conversation automation goes: phase 4, after the data and before agentic operations.
• Size context: Eurostat recorded AI use at 17% of small EU enterprises (10–49 staff) and 30.36% of medium ones (50–249) in 2025, against 55.03% of large ones.
• Governance reference: phase 2 maps onto the four functions of the NIST AI Risk Management Framework: Govern, Map, Measure, Manage.

What does an AI transformation roadmap look like for a company under 500 people?

A company under 500 people has no AI team and no programme office, so the roadmap is run by managers who already have full-time jobs. That rules out parallel workstreams: the phases run in sequence, each with one owner and a finish state someone can check. We call it the Baseline-First Roadmap, because nothing in it can be proved unless phase 1 is done.

*Elapsed times are our own planning ranges, not a published benchmark. The data range is shorter than the 6–18 months our explainer of what AI transformation means for a business gives for the whole data layer, because phase 3 cleans only the records use case one needs. That explainer defines the four layers this sequence moves through.

The Baseline-First Roadmap has one rule: a phase that has not reached its finish state does not hand over to the next one.

What should I do in week one of an AI transformation?

Measure, and buy nothing. In week one of an AI transformation, you do three things, and all three use systems you already have:

1. List every process where someone waits on a person. A customer waiting for a reply, a quote waiting for approval, an invoice waiting to be matched.
2. Write one number per process. Formula, current value, window. For example: median minutes from enquiry created to first contact attempt, last 90 days. If a process has no number, it cannot be in use case one.
3. Inventory the AI already in use. Ask each team which AI tools they use and what data they paste into them. This list becomes the input to phase 2.

Week one is done when one page lists the processes, their numbers and the tools in use. For customer-facing processes, the five-number AI readiness assessment checklist fills in that page in about 20 minutes with the CRM open.

Which phase of the AI roadmap takes the longest?

In an AI transformation roadmap, the data phase has the longest range of the gated phases, 1–6 months; phase 5 runs per workflow and phase 6 is ongoing by design. The data phase is also the phase with the least visible output, so it is the one that gets cut. It is slow because it means cleaning up years of manual entry, nobody usually owns it, and every later phase depends on it. Gartner reports that 63% of organisations either lack the right data management practices for AI or are unsure whether they have them. It predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data.

A firm under 500 people shortens this phase in one way only: it cleans the data for use case one, not the whole company. Before you start, write down the pass mark. Here is a worked example with illustrative inputs; substitute your own:

• Use case one needs 8,000 CRM contact records.
• Validation finds 25% fail at least one test (no valid mobile, no consent record, or a duplicate): 2,000 records.
• At 2 minutes per record for manual review, that is 4,000 minutes, or about 67 hours.
• At one person spending 8 hours a week on it alongside their normal work, that is 8–9 weeks. That is why this phase is measured in months, not days.

Where does conversation automation fit in an AI roadmap?

Conversation automation comes fourth: after the data, before agentic operations. Put it before the data phase and the agent works from bad records: opted-out contacts, wrong numbers, duplicates. Put it after agentic operations and you skip the cheapest proof point on the roadmap: one conversation type with a baseline, one owner, a weekly number and a human handoff at the booking.

Pick a conversation that already has a number: the first reply to new enquiries, follow-up on quotes, or reminders before an appointment. Change nothing else, then re-measure the same formula over a window of the same length. Four metrics decide it: speed to lead, contact rate (leads reached ÷ leads attempted), set rate (appointments booked ÷ leads reached) and show rate (appointments attended ÷ appointments booked). Write the scale, fix and stop thresholds down before launch, or the pilot cannot fail.

For revenue, use a slower number. LeadsNow’s methodology page defines sales lift as trailing three-month closed-deal revenue at month six, divided by the three months before launch. On that definition we report a 7x average across clients who supplied both figures; the same page discloses the median is closer to 4x, which is the number to plan with. Those figures come from LeadsNow’s own client work, which spans 50,769+ AI-booked sales appointments since 2017. Phase 4 is also the phase of this roadmap where AI appointment setting is the delivered service, not a method you build yourself.

When is my business ready for agentic AI?

A business is ready for agentic AI when its phase 4 metric has beaten baseline in two measurement windows in a row. At least one workflow must also have been redesigned, with a step switched off. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. It cites escalating costs, unclear business value and inadequate risk controls. Phase 1’s baseline is the answer to unclear value, and phase 2 is the answer to missing risk controls.

Phase 5 is the step most roadmaps skip. McKinsey’s 2025 survey found that AI high performers, about 6% of respondents, are nearly three times as likely as others to have fundamentally redesigned individual workflows. McKinsey says this redesign has “one of the strongest contributions to achieving meaningful business impact of all the factors tested”. Our notes on setting up agentic AI workflows cover the authority limits and review sampling phase 6 needs.

Which phase should my business start at?

Read down the table and stop at the first row that is true for your business. That row is your starting phase.

The 60% record-quality line and the 200-enquiry and 2,000-record volume lines are our own operating thresholds. They are not published research.

What it costs to run the roadmap in-house

Most of the cost of running an AI transformation roadmap in-house is staff time, not software. Phase 1 is a few days of an operations manager. Phase 2 needs someone who knows your privacy obligations, which may mean a lawyer or privacy officer rather than a vendor; this page is general information, not legal advice. Phase 3 is the review hours worked out above. Phase 4 needs an owned script, a reminder cadence for booked meetings and a weekly transcript review so the agent does not drift, which is the task teams drop first. Phase 5 is change management: whoever did the removed step needs new work. Phase 6 adds a reviewer.

Phase 4 is the phase that can be handed to a supplier and paid for on outcomes rather than hours. Our model is pay-per-result: 5–20% of the sales we help generate, or a fee per booked, qualified appointment. We do not charge a retainer or per seat. Phases 1, 2, 3 and 5 stay inside the business whichever way you go.

Frequently asked questions

How long does an AI transformation roadmap take for a mid-sized business?

Adding our planning ranges end to end, with the data phase scoped to one use case, a business reaches agentic operations between about month 4 and month 15; plan around month 9. That is 2–6 weeks for baseline and guardrails, 1–6 months for data, two conversation measurement windows of 2–8 weeks each, and 1–3 months for one workflow redesign. These are planning ranges, not a benchmark.

What should an AI roadmap template include?

For each phase: the first action, the finish state, an owner and an elapsed time. For each use case: the metric formula, its current value and the window it was measured over.

Should we hire a head of AI before starting?

Not for phases 1 to 4 in a business under 500 people. Those phases need a named owner with an existing metric, not a new role. McKinsey’s State of AI in 2025 found that about one-third of respondents said their organisations had begun to scale AI. Hire for scale once phase 5 has redesigned a workflow.

Why do AI roadmaps stall after the proof of concept?

They usually skip the baseline, the data or the business case. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. It gave four reasons: poor data quality, inadequate risk controls, escalating costs and unclear business value.

Do we need a governance framework before we start?

You need a one-page rule before customer data goes into any tool. A full framework is not required first. The NIST AI Risk Management Framework, released on 26 January 2023 for voluntary use, is a sensible structure to grow into.