The pilot works, then the real work starts

A lot of consulting firms can get an AI agent to do something useful in a pilot.

It might draft a proposal introduction. It might summarise a client interview. It might produce a first-pass market scan from public sources. The partner who sees it thinks, rightly, that there is something here.

Then the question changes.

Can it work across every service line? Can it use the firm’s best previous work without exposing one client’s information to another? Can it help a project team on Monday morning without someone manually uploading five folders and explaining the engagement context again?

That is where many agentic AI initiatives stop.

The problem usually isn’t the model. It isn’t even the agent workflow. The issue is fragmented context. Your firm has useful knowledge sitting across SharePoint, Google Drive, Teams, Notion, email, CRM records, proposal folders, project management tools, call transcripts, and the laptops of senior people. Each location has different access rules, naming conventions, and levels of reliability.

An agent can look impressive when someone gives it a clean folder and a tightly defined prompt. Scaling requires the agent to find, interpret, secure, and use information from the way your business actually operates.

For consulting and advisory firms in the $1 million to $25 million range, that gap has a direct cost. We commonly see annual leakage in the $80K to $300K range when senior people repeatedly research the same markets, recreate proposals, and spend hours hunting for work the firm has already paid to produce.

The opportunity isn’t to deploy an agent everywhere. It’s to build the context, controls, and operating process that let a small number of high-value agents do useful work repeatedly.

Why consulting firms have a context problem

Consulting is built on judgment, but much of the work around that judgment is repetitive.

A new opportunity arrives. A partner has an initial call, takes notes, and then asks a manager to find similar work, relevant case studies, bios, credentials, pricing references, and a proposal structure. The manager searches through old folders, messages a former project lead, pulls a few slides from a deck, and writes a draft. A major proposal can consume 20 to 40 hours before the firm even knows if it will win.

Then the engagement begins.

The team starts with secondary research. They review annual reports, competitor announcements, customer comments, regulatory material, industry reports, and internal notes. Some of this work is new. A lot of it is a version of research completed for another client six or twelve months earlier.

At the end of the engagement, the firm produces its best intellectual property. There are decks, financial models, interview summaries, workshop outputs, decision logs, and recommendations. Most of that material goes back into a project folder. It may be available in theory, but not usable at the moment another team needs it.

That is knowledge management debt. The firm pays for insight once to create it, then pays again because nobody can locate it, trust it, or safely reuse it.

AI agents expose this problem quickly. A Proposal Generation Agent cannot reliably pull relevant proof points if prior proposals are inconsistent, case studies have no structured metadata, and current pricing lives in a separate spreadsheet that only two people trust.

A Research Agent can’t build a credible starting brief if it has no way to distinguish public source material from a prior client’s confidential analysis.

A Knowledge Agent can’t answer a partner’s question across the firm’s project corpus if it cannot identify the client, industry, service line, date, document owner, confidentiality level, and source document behind the answer.

The first agent pilot often works because a capable person acts as the missing infrastructure. They curate documents, remove sensitive content, correct the output, and explain what matters. That support is reasonable in a test. It isn’t a scalable operating model.

The agent isn’t failing, the operating system is incomplete

When leaders say an AI agent “doesn’t scale,” they often mean one of four things.

First, it can’t access the right source material. The useful documents are scattered, permissions are unclear, or the content isn’t connected to the CRM opportunity or active project.

Second, it produces answers that are hard to verify. It may sound plausible, but a consulting partner needs to know where the claim came from, whether it is current, and if the source can be used in a client-facing deliverable.

Third, nobody has made decisions about client separation. A firm may want to reuse methodologies, public case studies, generic benchmarks, and approved credentials. It must not make Client A’s data available while serving Client B.

Fourth, the workflow stops with a draft. The agent produces output, but there is no defined reviewer, approval point, handoff, or measure of whether the work actually saved time.

These are business design problems. They need an owner. Treating them as a technology setup project creates more pilots and little adoption.

The firms that progress are usually more disciplined about scope. They don’t start by attempting a firm-wide AI knowledge brain. They choose one repeated workflow where the inputs, decisions, outputs, and review process can be made clear.

That might be proposal preparation for a particular service line. It might be engagement research in a specific sector. It might be making completed project assets searchable for internal use with strict client boundaries.

The underlying principle is straightforward. An agent needs enough context to act, but it must only access the context it is allowed to use.

Start with a client data and governance audit

Before expanding an agent pilot, map the data path it requires.

For a proposal workflow, ask where the opportunity data begins. It may start in HubSpot, Salesforce, a spreadsheet, or a partner’s inbox. Then identify where the agent would find approved case studies, past proposal sections, team bios, rate cards, credentials, and legal terms.

For an engagement research workflow, identify the trusted sources. Public web research may be appropriate. Internal sector points of view may be helpful. Prior client materials may be restricted completely, or only reusable after content has been redacted and approved.

For firm knowledge, identify what the firm wants to retain from every engagement, who classifies it, and when that classification happens. If people wait until the project closes, it often won’t happen. The better time may be when a document is saved, a meeting transcript is created, or a final deliverable is approved.

A practical audit should cover five areas.

Source systems. List where information lives today and which system is the source of truth. Don’t assume every document repository deserves to be connected. Old folders full of duplicates and unapproved working files can reduce answer quality.

Data structure. Identify the fields that make content usable. For consulting firms, that often includes client name, sector, geography, service line, engagement type, date, project status, confidentiality level, and approved reuse status.

Permissions. Define what different roles can see. A partner may need different access from a project manager. An agent should inherit those rules rather than operate as an unrestricted search tool.

Content controls. Decide what can be used for internal research, proposal drafting, client delivery, or model improvement. These are different categories. Put the rules in writing.

Human review. Define where a person checks the work. AI can prepare a proposal draft. The partner still owns the commercial claims, scope, pricing position, and client-specific recommendation.

This work doesn’t need a 60-page policy document. It needs decisions people can follow in the flow of work.

You can see the shape of this process in the AI audit for consulting firms. The goal is to identify the bottleneck that matters, the data required to fix it, and the controls needed to deploy safely.

What a production-ready proposal agent looks like

Take the Proposal Generation Agent as an example.

The trigger is a qualified opportunity in the CRM. The partner or business development lead provides a few inputs: client name, sector, problem statement, expected scope, timeline, likely stakeholders, and any points from the discovery call.

The agent then pulls from approved sources only. It identifies relevant past proposals based on service line and client situation. It retrieves approved case studies, credential slides, team bios, standard commercial language, and current pricing guidance. It doesn’t simply copy old content. It creates a structured first draft tied to the specific opportunity.

A useful output may include:

• A one-page opportunity summary based on the CRM and call notes
• A proposed problem statement and engagement objectives
• Two or three relevant credentials, with source links
• A draft workplan and timeline based on approved delivery patterns
• Open questions and assumptions that require partner input
• A pricing range or commercial framework for review, not automatic release

The partner reviews it. The delivery lead checks that the approach is viable. Commercial terms are approved by the person responsible for pricing. The final draft goes to the client-facing proposal process.

That is not an agent replacing a partner. It is removing the blank-page work and the document hunt.

The governance detail matters. The agent should cite the source of each case study and use only material labelled for approved reuse. It should flag a past client project as restricted instead of trying to summarise it. It should not turn a stale pricing spreadsheet into a client quote.

Once this is working, the firm can measure proposal cycle time, senior review hours, reuse of approved assets, and win rate by proposal type. The objective isn’t to make every proposal identical. It is to ensure expensive senior attention goes into the parts that require judgment.

If this is the constraint in your firm, Book a 60-min Omni Audit. In 60 minutes, we identify the priority workflow, map the data and governance gaps, and outline a practical deployment path. You get three outputs, not a sales deck.

Research and knowledge agents need different controls

The Research Agent is often a sensible starting point because the workflow has a clear deliverable.

At the start of an engagement, it receives the client, industry, geography, business question, and research boundaries. It searches approved public sources and selected internal material. It then produces sourced summaries, a structured industry view, key competitor signals, risks, and a one-page brief for the project team.

The team doesn’t accept the brief as truth. They use it to accelerate their first week. Every claim needs a source. Every source needs a date. Every uncertain finding needs to be labelled as a hypothesis rather than presented as a conclusion.

This is where a simple operating standard helps. Public research can be broadly accessible. Internal reusable points of view can be included when they are approved. Client-confidential source material stays separated unless explicit governance allows its use.

The Knowledge Agent has a broader job. It reads the decks, documents, and meeting transcripts the firm produces, then makes that corpus searchable through questions. A partner might ask, “What have we learned about margin improvement in industrial distribution over the past two years?” A project manager might ask, “Which workshop formats have we used successfully for operating model redesign?”

A useful answer should do more than generate a paragraph. It should provide the answer, identify the underlying documents, link back to the relevant source, and show the confidentiality status. If the question crosses a client boundary, the agent should return approved, anonymised knowledge or decline the request.

This is how the Omni Ops approach differs from a generic chatbot connected to a folder. The agent is designed around a real operational workflow, including inputs, permissions, outputs, review, and performance measures.

You don’t need to solve every knowledge problem before deploying one of these agents. You do need enough structure that the agent has a dependable job to do.

Fix the lowest-friction workflow first

Some consulting firms try to clean every repository before they automate anything. That can become a multi-year knowledge management programme with little business value along the way.

A better path is to select a workflow with four characteristics:

1. It occurs often enough to matter.
2. Senior people spend material time on it.
3. The required data can be identified and made available.
4. A human reviewer can validate the output quickly.

Proposal drafting often fits. Engagement research often fits. Creating reusable, classified project summaries at closeout can fit too.

Start with one service line or one segment of the firm. Build the data contract for that workflow. Decide the source systems. Establish permission rules. Define the output template. Run it with real work. Measure what changed.

Then expand from evidence, not enthusiasm.

The useful question isn’t, “Where can we use agents?” It is, “Where does fragmented context force our highest-cost people to repeat work?”

For some firms, the answer is cost of sale. For others, it is the first two weeks of every engagement. For many, it is the cumulative value lost when good work vanishes into disconnected project folders.

Our insights library and practical AI guides can help your team build shared language around these choices. But the work is specific to your systems, client commitments, and delivery model.

A practical checklist before you scale

Before moving a pilot into wider deployment, make sure you can answer these questions:

• What exact workflow does the agent own or support?
• What triggers the workflow?
• Which systems provide the source data?
• Which documents are approved for reuse?
• How are client confidentiality and access rights enforced?
• Can users see the source behind the agent’s output?
• Who reviews the work before it affects a proposal or client deliverable?
• What will you measure after 30, 60, and 90 days?
• What happens when the agent cannot find enough reliable context?

If your team needs a working document to get started, Deploy Your First Business Agent is a useful worksheet for defining the workflow, sources, approvals, and first success measure. You can also access the direct version here: download the business agent checklist.

The point is not to turn governance into bureaucracy. It is to prevent the predictable failure mode where the pilot relies on one enthusiastic person and can’t be trusted when other teams begin using it.

Move from isolated pilot to firm capability

A consulting firm doesn’t need dozens of agents to create value. It needs a handful that can work inside the firm’s real information environment.

The Proposal Generation Agent reduces blank-page proposal work. The Research Agent shortens the time between project kickoff and an informed team discussion. The Knowledge Agent makes more of the firm’s paid-for insight available without compromising client trust.

Each agent depends on the same foundation: clear source systems, usable metadata, client separation, permission controls, and human accountability.

If you can get that foundation right for one workflow, you have a repeatable way to scale. If you skip it, each new agent becomes another isolated demonstration that needs manual support.

For consulting firms carrying $80K to $300K in annual leakage from repeated research, proposal effort, and inaccessible IP, this is a commercial issue. Your senior people are expensive. Their time should go into client judgment, relationship building, and decisions that move the engagement forward.

See Omni for consulting firms to understand how we assess the workflow, data readiness, and deployment priorities. When you’re ready to put a real use case under the microscope, Book a 60-min Omni Audit. You’ll leave with three clear outputs and a practical next step, not a deck full of generic AI promises.