The pilot isn’t the hard part

Most consulting firms can get an AI agent to produce something useful in a demo.

Give it a past proposal, a client brief, a few case studies, and a prompt. It can produce a credible first draft. Ask it to summarize an industry, compare competitors, or search a folder of old reports. The output often looks impressive within an hour.

Then the pilot stalls.

The agent is not used for the next live proposal because the latest case studies sit in different drives. Pricing is stored in a partner’s spreadsheet. The sales lead has no agreed qualification brief to provide as an input. Nobody has authority to decide what goes into the approved draft, what needs human review, or who fixes an error.

That is why most agentic AI pilots never reach revenue workflows. The model may be capable. The operating workflow is not.

A recent Council Post on why agentic AI pilots fail to reach revenue workflows makes the same core point. The obstacle is rarely a lack of AI potential. It is the absence of a defined process, dependable source material, clear controls, and a person who owns the outcome.

For a consulting or advisory firm doing $1M to $25M in annual revenue, this matters because the cost is not abstract. Senior consultants are still writing proposals from scratch. Teams repeat the same secondary research during every new engagement. Valuable insight is buried in decks, meeting notes, spreadsheets, and documents that nobody can find when it counts.

Across firms of this size, the annual leakage from repeated work, slow cost of sale, and inaccessible knowledge can commonly sit in the $80K to $300K range. You won’t fix all of that with one agent. You can start recovering it when you fix one workflow well enough for an agent to operate inside it.

Start with a workflow that already matters

The wrong first question is, “What AI agent should we build?”

The better question is, “Which repeated workflow is costly enough to improve, stable enough to document, and important enough that someone will own it?”

For consulting firms, three workflows regularly meet that test.

Proposal and pitch production

A major proposal often consumes 20 to 40 hours of senior time before the work is even won. A partner shapes the angle. A manager hunts for relevant case studies. Someone updates credentials and bios. Another person tries to find the last pricing model used for a comparable engagement. The deck gets revised through email, chat, and document comments.

The win rate may be respectable. The cost of sale is still painful.

This is a good candidate for a Proposal Generation Agent (Omni ops), but only after you establish a repeatable proposal process. That means defining the opportunity brief, identifying approved content, setting pricing rules, and agreeing on who signs off before a draft reaches the client.

Engagement research and synthesis

Most firms say their work is bespoke. That is partly true. The client situation is unique, but much of the starting research is not.

Teams repeatedly compile market context, company background, competitors, recent announcements, operating model clues, and key financial information. They spend days finding and reading sources before they form an initial point of view. Across multiple clients in the same sector, the firm is often paying to discover the same baseline facts again.

A Research Agent (Omni ops) can create a structured starting brief with sources and summaries. It cannot replace senior judgment on the engagement hypothesis. It can remove a large amount of repetitive gathering and first-pass synthesis.

Knowledge capture and retrieval

Every engagement creates IP. It shows up in workshop notes, interview transcripts, slide decks, project plans, proposals, financial models, and final reports. Yet when the next team needs a relevant example, they ask around, search a shared drive, or rebuild the thinking.

That is knowledge management debt. The firm has paid for insight once, but cannot reliably put it to work a second time.

A Knowledge Agent (Omni ops) can read approved documents and transcripts, apply a usable structure, and answer questions across the firm corpus. But it needs a source-of-truth policy, document permissions, and a clear decision about what material belongs in the knowledge base.

These workflows don’t need to be perfect before you improve them. They do need to be sufficiently consistent that a person can explain how work moves from trigger to outcome.

Document the work before automating it

A workflow document does not need to be a 40-page process manual. For a first agent, one or two pages can be enough if it answers the right questions.

Start by mapping the current process as it actually happens, not as the firm says it happens.

For a proposal workflow, that might look like this:

1. A qualified opportunity reaches an agreed stage in the CRM.
2. The account lead completes a short opportunity brief.
3. The brief captures client objectives, scope, buyer roles, timeline, budget signals, and relevant credentials.
4. The proposal system retrieves approved case studies, team bios, service descriptions, and pricing guidance.
5. The agent produces a draft structure and first-pass narrative.
6. The proposal owner checks commercial positioning and factual accuracy.
7. A partner approves the final version.
8. The firm stores the completed proposal, outcome, and useful learnings in the knowledge base.

That is not complicated. It is specific.

Now compare it to the common version: a partner forwards a few emails, tells somebody to “pull together something strong,” and the team searches old folders. That is not a workflow an agent can reliably run. It is an informal coordination habit built around experienced people.

The distinction matters because agents need defined inputs, defined actions, defined handoffs, and defined exceptions.

If the work changes every time because nobody agrees on the process, the agent will feel unreliable. In reality, it is exposing a process that was unreliable before automation.

Clean the inputs that control the output

An agent doesn’t know which document is current unless your firm makes that clear.

This is where many pilots go wrong. The team connects a folder full of proposals and reports, sees that the agent can retrieve documents, and assumes it now has a usable knowledge system. It does not.

A folder can contain:

• Old service descriptions that no longer reflect the firm’s offer
• Case studies that are confidential or have expired client permissions
• Pricing tables from different years and different commercial models
• Multiple versions of the same deck
• Partner bios with outdated roles
• Research documents with weak or missing source links
• Client work that should never be available to a broad internal search

You do not need a perfect data estate to begin. You need a deliberately chosen, governed starting set.

For a proposal agent, create an approved content library. Include current firm credentials, approved service pages, selected case studies, current team bios, pricing principles, legal language, and a list of content that must never be reused without review.

For a research agent, define accepted source types and the expected output structure. For example, the agent may collect public filings, company websites, credible trade coverage, earnings materials, and selected databases your firm already licenses. It should cite its sources, flag uncertain claims, and distinguish fact from inference.

For a knowledge agent, decide what the firm wants to preserve from each project. A final deck alone may not be enough. The reusable value might be in the diagnostic framework, a workshop synthesis, a market map, a project retrospective, or an anonymized description of the result.

This is why our work in Omni ops begins with workflow and information design, not a rush to connect every tool the firm owns.

Assign an accountable owner, not a casual sponsor

Every production workflow needs an accountable owner. Not an executive sponsor who likes AI. Not the most technical consultant. The owner is the person responsible for whether the workflow produces a useful outcome each week.

For proposal generation, this might be the head of growth, commercial director, or a partner with responsibility for the sales process.

For research, it may be the practice lead or the person who runs engagement setup.

For knowledge management, it could be the operations lead, head of capability, or a senior consultant with a genuine mandate to improve reuse.

That owner should be able to answer five questions:

1. What triggers this workflow?
2. What inputs must be complete before the agent starts?
3. What output is considered usable?
4. When must a human review or override the agent?
5. What measure tells us the workflow is improving?

Without this ownership, the agent becomes a novelty. People try it when they remember. Nobody maintains the inputs. Exceptions accumulate. The pilot gets described as “promising but not ready.”

The owner does not need to manage prompts every day. They need to own the operating standard. That includes deciding when a source is outdated, when a workflow step changes, and when the team should stop using an output until an issue is fixed.

If you want a practical view of what that operating design looks like for your sector, see Omni for consulting firms. It focuses on the repetitive work that holds back delivery capacity and commercial throughput.

What a production-ready proposal agent looks like

A real Proposal Generation Agent is not a chat window where somebody types, “Write a proposal.”

It runs within a defined sequence.

First, the opportunity owner completes an intake brief. The brief is short enough that people will use it. It should capture the buyer’s problem, desired outcome, scope assumptions, decision process, deadline, relevant industries, and known objections.

Second, the agent reads that brief and retrieves only approved material. It finds case studies that match the sector and problem. It identifies relevant credentials. It applies current service language and commercial guardrails. It may prepare a set of clarification questions if the brief is incomplete.

Third, it creates a draft in the firm’s expected structure. That might include the client’s context, proposed approach, workplan, team, relevant evidence, pricing framework, assumptions, and next steps.

Fourth, a named reviewer checks the draft. The account lead reviews positioning. A partner reviews commercial judgment and client fit. The agent does not make promises, set final commercial terms, or invent evidence.

Fifth, the completed proposal is saved with meaningful tags. Opportunity type, industry, service line, outcome, and final pricing approach become useful inputs for future work.

That final step is often overlooked. Each completed workflow should improve the next one. If the firm never captures what was approved, changed, won, or lost, it keeps treating every proposal as a one-off.

The same principle applies to research. A Research Agent should receive a standard engagement setup brief, gather material against a defined checklist, produce a one-page briefing with citations, flag gaps, and store its output in a location that future teams can search. The partner then uses their judgment to form the point of view.

That is agentic work with guardrails. It is not outsourcing consulting judgment to a model.

Measure workflow performance, not prompt quality

A pilot should be assessed against business measures, not how impressed the team was by the first output.

For proposals, track:

• Senior hours spent from opportunity qualification to first client-ready draft
• Time to produce a first credible draft
• Percentage of proposal content pulled from approved assets
• Revision cycles before approval
• Cost of sale for major opportunities
• Win rate, while recognising that many factors affect it

For research, track time from project confirmation to initial briefing, source coverage, rework required by the engagement lead, and whether teams reuse prior research rather than restarting.

For knowledge, track how often people find a useful prior asset, time spent searching, and whether high-value project outputs are captured within a set period after closeout.

You do not need to pretend that every hour saved turns into immediate profit. Some capacity will be reinvested into better client work. That is still valuable. The key is to understand where senior time is going and whether the firm is reducing low-value repetition.

A practical first target is often to cut the time spent on a defined first-pass task by 25% to 50%, while keeping human review in place. The result will vary by workflow quality, source quality, and how much standardisation the firm is willing to accept.

Our resources and guides can help teams build baseline AI fluency. The more important work, though, is choosing the one workflow where better process design will create measurable capacity.

Don’t start with the most ambitious agent

The temptation is to build an all-purpose consulting agent. One system that researches, writes proposals, manages projects, captures knowledge, and answers every internal question.

That is a poor first deployment.

Start with a narrow workflow that has a clear boundary. Proposal first drafts are a strong option when the firm has recurring service lines and enough approved content. Research briefing works well when every engagement begins with a similar discovery phase. Knowledge retrieval can be valuable when the firm has a reasonably accessible body of past work and clear permission controls.

Then run the workflow with real work for 30 to 60 days. Keep a human in the review loop. Record failure points. Improve inputs and approval rules. Only then extend the agent’s scope.

The point is not to limit ambition. It is to build a reliable foundation before placing more client-facing or commercially sensitive work on top of it.

For a practical worksheet to select and prepare that first workflow, download Deploy Your First Business Agent. You can also access the direct worksheet here. Use it with your leadership team to identify the trigger, source data, owner, review points, and measures before you buy more software.

Turn AI interest into operating leverage

Consulting firms do not need another isolated AI experiment. They need operating leverage that protects senior capacity, reduces cost of sale, and makes hard-won knowledge reusable.

That begins with a workflow.

Choose one that repeats. Map the real process. Remove ambiguity from the inputs. Build a small approved source base. Put one person in charge of the outcome. Set review rules. Measure what changes.

Then the agent has a job it can actually perform.

An Omni Audit is built to find that starting point. In 60 minutes, we identify the workflows creating the most friction, estimate the value of improving them, and outline a practical first deployment. You get three outputs, not a generic slide deck: a workflow opportunity map, a prioritised agent plan, and a clear next-step recommendation.

Book a 60-min Omni Audit if you want to identify where your firm is losing capacity and which agent can reach production first.

You can also review the AI audit for consulting firms to see how we assess proposal work, research workflows, knowledge reuse, and the controls needed to make them dependable.

The best first agent is rarely the flashiest idea in the room. It is the one attached to a clean, owned workflow that people will use next week.

Book my Omni Audit and we will work through that workflow with you.