Service 03
Data and AI strategy
Most AI roadmaps fail on a data constraint that was discoverable in week one. We go looking for those constraints first, then sequence the work around what survives.
The problem this solves
The typical roadmap is a list of use cases ranked by value and effort, where effort was estimated without anyone opening the underlying data. It survives contact with reality for about a quarter.
The constraints that actually kill delivery are specific and findable: the corpus that turns out to be 60% scanned images at 150 dpi, the permission model that cannot be resolved per-user at query time, the contract that forbids the data leaving a jurisdiction, the process where nobody can define what a correct answer looks like.
A useful roadmap is one where each item has been checked against those constraints, and where the sequence is chosen so the first delivery de-risks the second.
Method
How we do it
Typically six to eight weeks, run with your team rather than delivered to them. The output is a plan your engineers helped write and can therefore defend.
01 · Opportunity capture
Structured interviews with the people doing the work, not only the people sponsoring it. The highest-value candidates are usually processes nobody has thought to describe because they are considered normal.
02 · Data feasibility
We open the corpora. Sampling for quality, coverage, permission model and revision hygiene, producing a feasibility rating per use case with the specific blocker named where one exists.
03 · Evaluability screening
For each candidate: what does correct look like, who can adjudicate it, and can we build a labelled set of a few hundred examples? A use case nobody can grade cannot be improved, and we would rather say so before it is funded.
04 · Sequencing
Ordered so early work builds reusable pipeline capability rather than isolated demos. The second use case should be cheaper than the first because of what the first left behind.
05 · Cost and capability modelling
Build and run costs per candidate, and an honest read on which parts your team can own now, which need hiring, and which are better bought.
06 · Governance
The approval path, the review points, and the model and data policies each use case has to satisfy, agreed with risk and legal during the engagement rather than discovered afterwards.
Deliverables
What you are left holding.
A sequenced roadmap
Twelve to twenty-four months, with dependencies, decision points and the explicit conditions under which an item should be dropped.
Feasibility findings per use case
Including the ones we recommend against, with the reason stated plainly.
A reference architecture
Target-state pipeline and platform design, sized to your estate rather than to a template.
A costed first delivery
Scoped tightly enough to start within weeks of the roadmap being signed off.
The hard parts
The questions worth asking us.
If a supplier cannot answer these specifically, they have not shipped one of these systems.
Will you tell us not to do something?
Regularly. The most common recommendation against is a use case where no one can define a correct answer, without that, there is nothing to evaluate, nothing to improve, and no way to know whether the system is getting better or worse. The second most common is a corpus whose permission model cannot be resolved per user, which makes the system undeployable to the audience that wanted it.
How is this different from a strategy deck?
We open the data. Feasibility ratings come from sampling real corpora, not from a workshop. It is the difference between a plan that says 'contract analysis: high value, medium effort' and one that says 'contract analysis: 12,000 agreements, 71% native text, 29% scanned pre-2016 at variable quality, no clause-level metadata, so budget for a clause segmentation stage and expect two weeks of labelling before anything can be measured'.
Related engagement
How this looks in practice.
Professional services
A two-year AI roadmap grounded in what the document estate could actually support
Fourteen candidate use cases screened against real corpora and against whether anyone could define a correct answer. Six survived; the sequence was chosen so the first delivery paid for the second.
Mid-market UK professional services firm, ~600 staff · 7 weeks
Read the engagementUsually engaged alongside
Start here
Talk to us about data & ai strategy.
Bring the estate, the constraint and the question you want answered. Forty-five minutes is usually enough to tell whether this is viable, and we would rather say so early.
- Prefer email
- hello@vectisflow.com
- Response time
- One working day, from a person who has read it.