Two to Four Hours After Every Meeting
Every prospective client of the agency starts with a discovery meeting. The meeting is recorded and transcribed, and the agency's services are described in an internal knowledge base. After the meeting a manager turns the conversation into the start of a deal: research the client and its market, size up competitors, pick the services that fit, sketch a first strategy and open tasks for the team.
All of this is done by hand and takes two to four hours per meeting, which caps how many prospects a team can pursue. The hardest part is knowing what to offer, because the match between service and problem sits with a few experienced people.
From Transcript to Draft
The assistant takes a meeting transcript and returns the groundwork of a deal as one structured result. The work is split into separate steps, each with its own input, output and record.
- Extract. A language model reads the transcript into a fixed schema: client, people, industry, size, needs, budget, tools and competitors named.
- Research. The client and a handful of its competitors are researched through web search and enrichment APIs. Every fact is stored with the address it came from, and a claim that cannot be tied to a source is flagged or dropped.
- Map. The client's needs are matched to services by retrieval over the agency's own knowledge base, each with the reason it fits.
- Draft. A first strategy: likely priorities, an order of work, a few quick wins and the questions still open.
- Review. The manager sees each claim beside its source and accepts, edits or rejects the draft. Only then does anything move downstream.
The output keeps the same shape every time, in the language the meeting was held in. The assistant is designed to run in scheduled batches. If an external source is down, its block is skipped and flagged, and the rest of the run completes.
Where a General Chatbot Falls Short
- A general model invents facts. Asked to profile a company and its competitors, it returns fluent text with made-up claims. A made-up competitor reads exactly like a real one.
- One prompt hides the work. With transcript in and proposal out, the manager cannot see what is grounded and what is guessed, or fix the research before it feeds the strategy.
- Services picked from memory. The model recommends services the agency does not sell and misses the one that fits.
- Fluency hides fabrication. An invented fact is written in the same assured tone as a sourced one. The most reliable test is whether the claim ties to a real, retrievable source.
- Full automation with no checkpoint. In this agency's work, a wrong brief wired straight to tasks becomes wrong work on something a client will see.
Technology Stack
Sources and Review
The preparation ends up in front of a client, so the agency's name is on every claim. The assistant prepares the material; the manager decides what leaves the building.
- Every claim carries its source. Research a manager cannot check is research a manager cannot forward.
- The recommendation comes from the catalog. The assistant proposes only services the agency offers, read from the knowledge base rather than recalled.
- Steps stay inspectable. A good step is kept when a later one is wrong, and every run is logged end to end.
What the Assistant Is Specified to Do
The figures below come from the assistant's specification.
- Two to four hours of manual preparation per meeting, which the assistant is designed to replace with a researched draft.
- No claim reaches the draft as fact without a source.
- No draft passed on without a manager's review.
- Recommendations drawn only from the services the agency offers.
- A run that completes, with the gap flagged, when an external source is unavailable.
Similar Preparation Work Elsewhere
Skilled people turn a recorded conversation into researched, client-facing preparation in many businesses, usually under the same conditions:
- the preparation follows the same shape after every meeting;
- knowing what to offer sits with a few experienced people;
- the result is shown to a client, so an unchecked error is expensive;
- the recording or transcript already exists.
Presale in services firms, consulting proposals, deal screening and tender responses are close relatives. The same steps can start from emails, call notes or a CRM record instead of a recording, and write into the client's own CRM or project tools. Our glossary covers the core techniques in retrieval-augmented generation and human-in-the-loop, and the build itself falls under enterprise knowledge assistants and generative AI development.
of manual preparation per meeting the assistant is designed to take over
of claims in a draft designed to carry a link to their source
separate steps from transcript to draft, each with its own record