A roadmap is the cheapest document a vendor can deliver and one of the most expensive to get wrong. It can be produced in a week, look convincing in a board pack, and still leave the business unable to say which process to automate first, who owns the data, or what happens when the pilot fails. The purchase decision is therefore not "who writes the best report" but "whose report will survive contact with operations".
This guide is for an SME owner in Bangladesh who has decided to adopt AI deliberately rather than through a series of tool purchases. It covers the brief, provider types, a scorecard, pricing questions, pilot design and exit terms.
Start from the decision, not the technology
A useful roadmap answers four questions: which processes are worth changing, what must be true before each change is safe, in what order the work should happen, and how much the business is willing to spend before seeing a return. If a proposal cannot answer those in plain language, it is a technology survey with a cover page.
Write down the decision you are trying to make. "Choose three processes to automate over the next four quarters, with a budget ceiling and a named owner for each" is a decision. "Become AI-ready" is not. The narrower statement makes it possible to judge the roadmap against something other than its own ambition.
Requirements to put in the brief
- Current-state evidence: process maps, volumes, cycle times, error rates and the systems that hold the data. Ask how the vendor will collect this, and from whom.
- Use-case shortlist with rejection reasons: not just the opportunities but the ones ruled out, and why. A roadmap that only adds is a sales list.
- Data readiness per use case: which fields, which source system, who owns quality, and what the minimum viable dataset looks like.
- Integration map: the identity provider, ERP or accounting system, CRM, communication channels and any partner APIs.
- Governance and risk: data classification, retention, vendor access, human review for consequential decisions, and a named accountable owner.
- Sequencing with dependencies: what must be fixed before the next step, and which work can run in parallel.
- Cost envelope and business case: licence, implementation, integration, change management and ongoing operation, separated rather than bundled.
- Success measures and stop conditions: the observable result that justifies continuing, and the result that should end the initiative.
Four kinds of provider you will meet
Strategy consultancies are strong on framing, process analysis and executive communication. Their weakness is delivery: the roadmap may be sound, but the firm does not build the workflow, so scope and cost reappear at handover.
Automation and software vendors understand implementation and usually have reusable components. Their bias is toward the work they already know how to deliver, which can push a data problem into a tool project.
AI platform resellers bring licences, model access and vendor roadmaps. They are useful when the constraint is capability, but a licence is not a plan. Ask what happens to prompts, agents and integrations if the reseller relationship ends.
An internal lead with a fractional adviser keeps context and costs less in fees, but depends on spare capacity. It works when someone senior can own the roadmap for a quarter and the adviser is used for challenge rather than production.
Most SMEs end up with a combination. Know which bias you are buying, and write the other roles into the plan.
A scorecard for comparing proposals
Score each proposal from 1 to 5 against the same evidence. Do not average away a weak data plan because the presentation was good.
| Criterion | Evidence to request | Score 1–5 |
|---|---|---|
| Process evidence | Interview notes, volumes and a baseline for each shortlisted process | |
| Data plan | Named fields, source systems, owners and a minimum viable dataset per use case | |
| Sequencing | Dependencies, prerequisites and a first step that can be completed in weeks | |
| Governance | Data classification, human review, vendor access and retention rules | |
| Delivery capability | Who builds, who tests, and whether the same team supports the result | |
| Cost transparency | Separate lines for discovery, build, integration, licences and support | |
| Handover | Documentation, training and a named person for transition support |
Pricing questions that expose scope
- Is discovery a fixed fee with a defined deliverable, or an hourly engagement with no ceiling?
- How many workshops, interviews and site visits are included, and who attends from our side?
- Is the roadmap priced per use case, per department or per outcome? What happens if we add a fourth process?
- Are software licences passed through at cost, or bundled into the fee? What is the renewal basis?
- Is data cleanup and integration quoted separately? If not, which assumptions make it "included"?
- What does a failed pilot cost us, and what is delivered if we stop at the end of it?
- Who owns the prompts, agent configuration and documentation produced during the engagement?
Run a paid pilot before signing the full roadmap
Pick one workflow with a visible output and a small number of users. Run it for four to six weeks with a baseline, a written acceptance test and a weekly review. The pilot should test the vendor's working style as much as the technology: how they handle a data gap, a user who refuses the new process, and a result that misses the target.
Define the kill criteria in advance. A pilot that cannot fail is a demonstration with invoices attached. If the vendor resists a written acceptance test, that is information about the main engagement.
Exit terms and knowledge transfer
Ask for the plan in an editable format, the source data and mappings, and a session that walks your team through the assumptions. If the roadmap depends on a proprietary assessment tool, confirm that you can keep the outputs after the contract ends. Transition assistance should have a stated number of days and a named person, not a promise of goodwill.
Red flags
- A roadmap that names tools before it names processes.
- Savings or productivity claims with no baseline and no method.
- No mention of data ownership, retention or vendor access.
- A pilot that is free but has no acceptance criteria.
- One person holds all the knowledge, and no documentation is in scope.
- A timeline that assumes every department will change at the same speed.
Decision framework
Choose a strategy-led engagement when the problem is disagreement about priorities and the build partner is already known. Choose a builder-led engagement when the process is understood and the risk is delivery. Choose an internal roadmap with a fractional adviser when you have a credible owner and need challenge rather than capacity. Keep the first commitment small enough to reverse.
We deliver AI adoption and agent work under Business Automation. The pricing page explains how discovery, build and retainer work are separated. For the longer view, read how AI adoption in SMEs changes year by year.
Sources
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- ISO, ISO/IEC 42001: Artificial intelligence management system, accessed 17 November 2026.
- Gartner press release, Applying uniform governance across AI agents will lead to enterprise AI agent failure, accessed 17 November 2026.
- EU AI Act, Implementation timeline, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.