Partners and senior leaders have every reason to be wary of the noise around AI. Each week brings a new tool, another vendor promise and another headline about jobs, ethics or compliance. Meanwhile, the core systems are still creaking under what you ask of them on an ordinary Tuesday.

That contrast points to a practical starting place. Before choosing an AI product, map one commercially important workflow as it operates today. The exercise exposes whether the firm has a model problem, a process problem, an integration problem or some combination of the three.

This article is deliberately narrower than our broader guide to AI adoption. It explains how to choose and map the workflow that a pilot would enter, so the leadership, technology and risk teams can make a grounded decision before a licence or build commitment.

Choose a workflow, not an abstract use case

“Use AI for proposals” is too broad to design or govern. A proposal may involve opportunity qualification, research, finding relevant credentials, choosing a team, drafting, pricing, conflict checks, partner review and submission. Each step has different data, owners and consequences.

Define the part you want to examine with a clear beginning and end. For example: from an approved opportunity brief to a first proposal draft ready for partner review. Name the person who receives the output and the decision they will make with it.

Then test whether the workflow matters enough. Record its volume, elapsed time, hands-on effort, rework and effect on clients or revenue. A task that irritates three people twice a year may be a poor pilot even when the demonstration looks impressive. A repeated bottleneck that delays high-value work gives the team a clearer baseline and reason to act.

Map what actually happens

Process diagrams often describe the approved route. A useful AI assessment needs the real one, including spreadsheets, copied data, personal templates and conversations that never reach the system.

Bring together the people who request, perform, review and support the work. Walk through a recent example and capture:

  • The trigger that starts the process.
  • Every input and where it comes from.
  • Each decision, approval and hand-off.
  • The systems and documents used at each step.
  • Common exceptions and how people recognise them.
  • The output, recipient and definition of acceptable quality.
  • The record retained after the work is complete.

Ask where people wait, rekey information, search for material or correct earlier work. Those points may be suitable for automation, although they may also reveal a simpler process or integration fix.

A proposal walkthrough may show that partners use different qualification criteria, credentials live in several locations and the final price is approved through an email chain. Adding generation at the drafting step would accelerate one part of an inconsistent process. The first work in that situation is agreeing the brief, evidence source and approval route.

Inspect the data at each step

For every input, identify its owner, format, sensitivity, source and expected quality. Check whether the proposed system can access it through an approved route and whether that access is limited to what the task requires.

Sample the material rather than relying on descriptions. Open recent records. Look for duplicates, missing fields, obsolete templates, inconsistent labels and documents that people know to ignore. If experienced staff apply unwritten judgement to distinguish reliable material, capture that rule or retain the human decision.

This is where AI turns the lights on. A system that retrieves from a neglected knowledge base can make existing disorder more visible and more persuasive. Cleaning every repository may be unnecessary for a bounded pilot. Define an approved source set, exclude weak material and make gaps visible to the reviewer.

Mark decisions and exceptions

Use the map to separate four kinds of activity:

  1. Move or format: transfer data, convert a document or populate an approved template.
  2. Find or summarise: retrieve relevant material and produce a reviewable synthesis.
  3. Recommend: rank options, flag a risk or propose a next action.
  4. Decide or act: approve work, communicate externally or change a client, financial or regulatory outcome.

The category shows where consequence can increase; the risk assessment determines the control. A drafting assistant working from approved material is different from a system that selects which credentials to claim or sends the proposal without review.

List exceptions explicitly. What happens when a required field is absent, sources disagree, the matter belongs to a restricted client or the output falls outside the system's intended scope? Define when the workflow stops, who receives the case and what information they need. Compliance can govern a visible hand-off more confidently than a promise that the model will handle ambiguity.

Decide whether AI is the right intervention

The map may show that AI is unnecessary. A standard form could improve the input. An API could remove rekeying. Clear ownership could shorten an approval. A searchable, maintained repository could solve the retrieval problem with less uncertainty.

Use AI where its capability addresses the measured constraint and a test can demonstrate improvement. Define the comparison before the pilot: time to a reviewable draft, corrections per output, source coverage, exception rate, user effort or another measure tied to the workflow.

Also define the minimum acceptable quality and stopping conditions. If the system cites material outside the approved source set, exposes restricted information or increases review effort, pause and investigate. A pilot exists to discover those limits before the workflow becomes business as usual.

Give compliance something legible to govern

Risk and compliance functions struggle when a proposal consists of a product name and a broad promise. Give them the workflow map, data inventory, proposed permissions, accountable owners, tests, exception route and monitoring plan.

The useful question becomes specific: can this system assist this step, with this information, under these controls? The answer may be yes, yes with conditions, or no. Each response gives the project team something it can act on.

At the next leadership meeting, ask whether the firm could show where the pilot's data comes from, where it goes, who checks the output and who remains accountable. If those answers are missing, completing the workflow map is the project that comes first.

An external review can be useful when internal sponsors are already committed to a tool. Its first job should be to test whether the investment belongs in AI, integration, data or process design, with permission to recommend any of the four.