Agent costs don’t fall the way chat costs do

A chat assistant is usually easy to price. Someone asks a question, gets an answer, and the session ends. The cost is tied mainly to the volume of input and output.

An AI agent is different. It can plan a task, search internal files, call external sources, compare results, run validation steps, store a summary for later, then retry parts of the workflow if the output doesn’t meet a rule. Each step may involve a different model, a new prompt, another document retrieval, or a tool call.

That matters for consulting and advisory firms.

A partner may see an agent produce a credible first proposal draft in 15 minutes and assume the economics will improve as more people use it. But an agent that has to inspect 30 past proposals, find relevant case studies, check a rate card, validate client facts, and refine its own work can cost far more than a simple chat prompt. If it gets used across every opportunity without limits, the monthly bill can become hard to explain.

Recent reporting on Gartner’s view of agentic AI has brought this issue into focus. More usage doesn’t automatically create the economies of scale business owners expect. In many agent workflows, greater volume means more reasoning runs, more context, more retries, and more orchestration.

The answer isn’t to avoid agents. The answer is to price the workflow before you deploy it.

For a consulting firm doing $1 million to $25 million in revenue, that means knowing three things before a pilot becomes a company-wide tool:

1. What manual activity the agent replaces or reduces.
2. What each completed workflow is allowed to cost.
3. What quality threshold must be met before the work reaches a client or senior reviewer.

This is a practical operating issue, not an AI theory discussion. Done well, it helps protect the $80K to $300K annual leakage band we often see across proposal effort, repeated research, and knowledge that can’t be reused.

Start with the expensive work, not the impressive demo

Most consulting firms don’t have a shortage of AI ideas. They have a shortage of clarity around where time is actually being spent.

Take a major proposal. A partner takes the first call. An associate digs through prior decks. Someone searches shared drives for relevant credentials. Another person tries to find the right case study and confirm what can be said publicly. Pricing gets checked against an old spreadsheet, then changed after a partner review. The final narrative comes together late because the team is also delivering client work.

A substantial proposal can absorb 20 to 40 hours before the first client-ready version exists. A good win rate doesn’t make that cost disappear. It just makes it easier to tolerate until utilisation drops or the pipeline gets busy.

Research work has a similar pattern. At the start of a strategy, transformation, or market-entry engagement, a team spends days collecting public market data, company information, competitors, news, filings, and prior internal points of view. Some of that work is specific to the client. A surprising amount isn’t.

Then there is the knowledge management problem. Each project produces decks, interview notes, workshop outputs, transcripts, models, and final recommendations. Much of that intellectual property is stored in a place no one searches when the next engagement begins. The firm pays for the same insight twice, often with more senior time than the first time.

Those are useful places for agents because the work is structured enough to define a workflow and expensive enough to justify governance.

The mistake is building an agent around a vague instruction like, “make our proposal process faster.” That creates a tool no one can price.

A better instruction is specific:

Produce a first-draft pursuit brief using approved firm material, retrieve no more than 12 source documents, flag any claims without evidence, and stop if the workflow cost reaches the set ceiling.

That gives the team something to measure.

What a priced proposal workflow looks like

Our Proposal Generation Agent is designed for a workflow like this. It doesn’t replace the partner’s judgement on scope, positioning, or pricing. It handles the repeated assembly and first-draft work that consumes time before that judgement can happen.

A typical workflow might run in seven stages.

1. Intake and opportunity classification

The agent receives the meeting notes, client website, opportunity stage, proposed service line, geography, estimated deal value, and any stated buying criteria.

It classifies the opportunity before it starts retrieving documents. A turnaround project should not pull the same material as a procurement transformation or a leadership advisory engagement.

This first step needs a small, controlled input set. If people paste every email thread and every meeting transcript into the workflow, costs rise before the agent has done anything useful.

2. Retrieve approved firm evidence

The agent searches a curated knowledge set for relevant case studies, biographies, methodologies, credentials, and prior proposal language.

The word “approved” is important. If the retrieval library includes old drafts, unverified claims, or commercial documents with expired pricing, the agent may create a polished proposal with the wrong facts.

A sensible retrieval limit might be 8 to 15 documents for a standard opportunity, with a separate escalation path for complex bids. The agent should return the sources it used, not just the draft it produced.

3. Build the proposal structure

The agent prepares an outline based on the opportunity type. It might include the client’s situation, relevant point of view, approach, team, scope assumptions, timeline, commercial structure, and proof points.

At this stage, it should not invent client pain points or claim results the firm can’t support. A good agent marks gaps clearly, such as “partner input required” or “no approved case study found.”

That behaviour is valuable. It stops the team from treating fluent wording as verified content.

4. Draft the narrative

The agent creates a tailored first draft using the retrieved source material and the agreed structure. This is where token use often rises. Long context windows, lengthy proposal templates, and repeated rewriting can turn a low-cost workflow into an expensive one.

The firm needs a rule for when drafting stops. For example, one full draft, one self-review pass, and one revision based on defined checks. Not unlimited rewriting until the agent decides it sounds better.

5. Validate facts, commercial references, and claims

A separate validation step checks the draft against the source set. It should identify unsupported claims, missing source links, confidential client references, outdated team biographies, and price language that doesn’t match the approved rate card.

This is not the place to save money blindly. Validation is one of the workflow steps that protects margin and reputation. But it still needs a budget. A workflow that rechecks every sentence against a large document collection may be unnecessary for a $30,000 opportunity.

6. Hand off to a human reviewer

The output goes to the partner or bid lead with a short evidence pack. They can see what the agent used, what it couldn’t find, what it flagged, and what assumptions need a decision.

The agent speeds the path to judgement. It shouldn’t create the illusion that judgement is no longer required.

7. Log cost and outcome

For each completed proposal workflow, record the model cost, retrieval volume, number of tool calls, elapsed time, human review time, and final outcome.

That last part matters. If an agent costs $12 per proposal and saves six hours of associate time, the case may be clear. If it costs $90, creates unreliable claims, and still needs four hours of cleanup, it needs redesigning.

The goal is not the cheapest possible workflow. The goal is an economically sound workflow that produces a reliable first draft.

If you want help mapping that structure against your own pursuit process, Book a 60-min Omni Audit. We use the session to identify the highest-value workflow, define the inputs and controls, and put a practical number against the opportunity.

Set cost ceilings before people get access

The most useful control is a per-workflow ceiling.

Don’t begin with a broad monthly AI budget. That won’t tell you which process is consuming spend or whether it creates enough value. Begin with a ceiling for each completed unit of work.

For a consulting firm, the unit might be:

• One qualified proposal draft
• One research brief for an active engagement
• One knowledge search with cited evidence
• One client meeting preparation pack
• One project closeout knowledge capture

The ceiling should be connected to the commercial value and manual effort of that unit.

A high-value, complex proposal may justify a higher ceiling because it replaces meaningful research and coordination. A basic meeting summary should have a much tighter limit. It doesn’t need five reasoning loops, 40 document reads, and persistent memory to produce a useful output.

There are four main areas to meter.

Tokens and context

Track input and output tokens separately. Input costs can rise quickly when the workflow repeatedly sends old proposals, long transcripts, and large knowledge extracts to the model.

A common fix is to retrieve narrower extracts rather than passing full documents. The agent needs the relevant paragraph about a case study, not a 60-page deck every time.

Tool calls and searches

Every search, browser action, document conversion, database call, or external research request can have a cost or a latency impact. Put limits on how many sources the agent can inspect before it returns a gap for human review.

For the Research Agent, this could mean a defined research plan, a capped number of external sources per topic, and an evidence table that shows what was found. The aim is not to scrape the internet without restraint. It’s to produce a structured one-page brief that gives the engagement team a starting point.

Reasoning and retries

Some workflows use multiple internal passes to plan, critique, and improve an answer. Those passes can improve quality, but they need a purpose.

Define which checks are required. Perhaps the proposal agent runs one claim-validation pass and one brand or tone check. It shouldn’t retry endlessly because a broad prompt tells it to produce the “best possible” answer.

Memory and retained context

Memory can make an agent more useful over time, particularly where it needs to understand a firm’s service lines, methods, and preferences. It can also create cost, privacy, and data quality issues.

Retain facts that have a clear future use. Don’t store every draft thought, client conversation, or intermediate output as permanent memory. That creates a larger and less reliable context pool for future workflows.

Build a cost guardrail and a quality guardrail

A ceiling on cost without a quality standard encourages bad shortcuts. A quality standard without a ceiling encourages expensive automation. You need both.

For each agent, define a small scorecard before rollout.

Measure	Example rule
Workflow cost	Stop or escalate when the ceiling is reached
Source coverage	Every factual claim must link to an approved source
Human acceptance	Reviewer can use the output with minor edits
Completion time	Deliver the first draft within an agreed window
Escalation rate	Flag unclear scope, pricing, or unsupported claims

Run the pilot on 10 to 20 real workflows before offering it to the whole firm. That is enough volume to find the expensive edge cases without locking the business into a poorly designed process.

Review the outliers. The valuable learning is often in the workflow that costs four times more than expected. Did it retrieve too many documents? Did a long transcript get passed repeatedly? Did the agent retry because the request wasn’t specific? Did it need an external research task that should have been a separate workflow?

Those answers help you redesign the work rather than simply choosing a cheaper model.

Research and knowledge agents need the same discipline

Proposal creation is only one use case. The same economic discipline applies to research and knowledge management.

A Research Agent can run a defined engagement-start workflow. It takes a client name, industry, geography, stated problem, and priority questions. It collects approved public sources, creates short summaries, records citations, highlights assumptions, and produces a one-page brief for the team.

The cost ceiling should reflect the stage of the engagement. A pre-sale briefing may get a smaller budget than a funded diagnostic. If the engagement requires deep market research, that should be deliberately scoped, not hidden in the agent’s background activity.

The Knowledge Agent has another job. It reads selected decks, documents, and meeting transcripts, indexes the approved content, and answers questions across the firm’s corpus. Done well, it reduces the time spent asking, “Has anyone done this before?”

But it must have boundaries around ingestion. Not every draft belongs in the knowledge base. Not every client document can be shared across teams. The workflow needs access controls, retention rules, source citations, and a process for removing content that shouldn’t be there.

That is the difference between a useful knowledge asset and a large folder with a chat box on top.

For broader operating examples, the Omni operations approach is built around defined business workflows, not generic AI access. You can also find practical implementation material in our AI guides when you are deciding where an agent should start and where it should stop.

Find the leakage before you fund the build

For firms in this size range, the $80K to $300K leakage band isn’t usually one dramatic failure. It is accumulated senior time, duplicated analyst work, slow proposal production, and knowledge that never gets reused.

The financial case should include more than software costs.

Look at:

• Hours spent building proposals that don’t require original thinking
• Repeated research tasks across similar client assignments
• Senior review time spent correcting basic facts and formatting
• Time lost locating past examples and credentials
• Delays that reduce the number of opportunities the firm can pursue
• Rework caused by using outdated case studies, scope language, or pricing

Then compare those costs with a controlled agent workflow. If the workflow doesn’t create a clear reduction in effort, improved turnaround time, or better reuse of IP, don’t scale it.

A useful first step is the AI audit for consulting firms. It helps identify which workflows have enough volume, repeatability, and commercial value to justify an agent, and which ones should remain human-led.

Use a worksheet before you build

If you are preparing an internal pilot, download Deploy Your First Business Agent as a working checklist. It helps you define the job the agent owns, the inputs it can use, the decision points that require a person, and the measures that prove it is earning its place.

You can access the direct worksheet here: Deploy Your First Business Agent.

Use it with the team that does the work, not just the person sponsoring AI. Associates, engagement managers, operations leads, and partners will see different failure points in the workflow. Those details determine the real cost.

Deploy a controlled workflow, not an open-ended agent

The best first agent is rarely the one that promises to run the firm. It is the one with a clear start point, a defined output, approved data, a quality check, and a cost ceiling.

For many consulting firms, that is a Proposal Generation Agent or Research Agent. Both address expensive repeated work. Both can be tested on live activity. Both create a visible before-and-after comparison for time, quality, and cost.

Start with one workflow. Meter it from the first run. Review the exceptions. Improve the source library. Tighten the instructions. Keep the human approval point where commercial judgement matters.

Then decide if it should scale.

If you want a 60-minute working session on where agent costs could get out of control in your firm, Book a 60-min Omni Audit. You’ll leave with three practical outputs: a prioritised workflow opportunity, a first-pass leakage estimate, and a deployment plan. No deck, no generic AI roadmap.

You can also see Omni for consulting firms to understand how we map proposal, research, and knowledge workflows into an operating model that your partners can actually govern.