The cheap AI pilot can become an expensive operating habit

A consulting partner tries a generative AI tool for a proposal. They paste in a short brief, ask for a first draft, make a few edits, and get a useful result in 90 seconds.

The cost looks trivial.

Then the firm decides to build an agent that can do the full job. It needs to read the opportunity brief, search past proposals, pull relevant case studies, identify sector language, check pricing assumptions, create a draft, critique the draft, revise it, and prepare an executive summary.

That is no longer one prompt. It is a chain of model calls, document retrievals, tool calls, revisions, and checks. Each step consumes tokens. Some steps consume the full context from prior steps again. If the agent loops because the source material is unclear, the bill rises while nobody is watching.

This is the issue behind recent Gartner commentary reported by Computer Weekly. Agentic AI does not necessarily get cheaper just because you run more of it. In many cases, scale exposes the full cost of multi-step workflows.

For consulting and advisory firms, that matters because your work is document-heavy, research-heavy, and high-context. A useful agent may need access to dozens of slides, prior statements of work, client notes, transcripts, spreadsheets, and external sources. The potential value is real. So is the risk of treating token usage as someone else’s technical problem.

Firms in the $1M to $25M range commonly have enough repeatable work to gain from agents, but not enough spare margin to absorb an uncontrolled AI bill. Across a year, the broader leakage opportunity we see in consulting businesses is often in the $80K to $300K range. That includes senior time spent rebuilding proposals, repeating research, and hunting for knowledge that already exists inside the firm.

The answer is not to avoid agentic AI. It is to cost each business task before you put it into a recurring workflow.

Why multi-step agents use more tokens than simple prompts

A standard AI interaction is usually straightforward. You provide instructions and source material, the model produces an answer, and the interaction ends.

An agentic workflow works differently. It is designed to decide what to do next. It might search a folder, inspect five documents, find missing information, run another search, draft an answer, review that answer against a checklist, then revise it.

Every stage has a token cost.

Tokens are the chunks of text a language model reads and writes. A slide deck, a transcript, a proposal, and an email thread can each add thousands of tokens. The model is billed based on what it processes and what it produces, although exact pricing depends on the model and provider.

The cost rises in a few predictable ways.

Large context is repeatedly passed through the workflow. If an agent reads a 40-page proposal and carries that content into six later steps, the same information may be processed repeatedly. This is often the biggest hidden issue in early prototypes.

The workflow branches. A research agent might investigate three competitors, five market themes, and four client priorities. That can quickly become 60 separate sub-tasks before it writes the final brief.

Agents critique and retry their own work. Review loops can improve quality. They can also double or triple usage if the instructions are vague, the quality standard is unclear, or the source data is poor.

Tool results add more content. Web searches, document retrieval, CRM records, meeting transcripts, and spreadsheets all expand the context window. An agent that can access everything is not automatically an agent that should read everything.

Human approval arrives too late. If the agent has already completed 18 steps before a manager sees a weak assumption, it has already incurred the cost of that work.

A simple illustration makes the point. Imagine a one-step prompt that processes 3,000 tokens of source material and produces 800 tokens. Now compare it with an agent that runs eight stages. It retrieves documents, summarizes them, creates an outline, drafts sections, checks each section, and prepares a final version.

Even if each individual step appears affordable, the full task may process 30,000 to 100,000 tokens once you include repeated context, retrieval results, and review passes. Add more research branches or longer client history and the number can rise quickly.

The issue is not that agents are inherently wasteful. The issue is that a workflow must be designed around a business outcome, a fixed scope, and a measurable unit cost.

Start with the manual work that is already costing you

Token costs are only one side of the equation. A consulting firm should compare the AI cost against the time it removes, the quality it improves, and the risk it reduces.

The best first use cases are not broad ambitions like “build an AI consultant.” They are contained pieces of work that happen repeatedly and have a visible cost.

Proposal work is often the first pressure point

A major proposal can consume 20 to 40 hours of senior and mid-level time. The partner frames the opportunity. A manager searches old decks. Someone rewrites a case study. Finance checks the commercial model. The team adjusts language for the client’s industry, then changes it again after a partner review.

The win rate might be acceptable. The cost of sale is what hurts.

A Proposal Generation Agent from Omni ops can handle a structured part of this workflow. It starts with an opportunity brief containing the client, sector, scope, deal stage, likely team, and commercial constraints. It retrieves approved past proposals, relevant case studies, standard capability language, and current pricing guidance. It then creates a tailored draft with citations back to the internal source material.

The agent should not be allowed to browse every document the firm has ever produced. That is how token spend and errors both grow. It needs a curated retrieval set, current templates, approved examples, and clear limits on the number of documents it can review for a single opportunity.

A good workflow might look like this:

1. A manager completes a five-minute opportunity intake.
2. The agent classifies the opportunity by service line, sector, and likely scope.
3. It retrieves up to five approved proposals and three case studies.
4. It creates an outline before writing full copy.
5. A human approves the outline.
6. The agent drafts the proposal using the approved structure.
7. A partner reviews commercial terms, claims, and positioning.

That human gate after the outline is important. It stops the agent from spending tokens drafting 25 pages around the wrong story.

Research gets repeated across engagements

Many consulting firms begin an engagement with a familiar scramble. An analyst gathers annual reports, investor presentations, news coverage, market reports, competitor information, and interview notes. A manager turns it into slides. The team discovers that another practice area did similar work for a different client six months earlier, but no one can locate it.

This repeated work compounds across the firm.

A Research Agent can run a structured research process at the start of every engagement. It can collect approved external sources, summarize company and industry developments, identify open questions, and produce a one-page brief with links back to the evidence.

The useful version is not an agent that generates generic industry commentary. It is one that operates from a defined research plan.

For example, a market-entry engagement might require:

• Company position and revenue model
• Customer segments and buying triggers
• Five direct competitors
• Recent regulatory changes
• Relevant acquisitions or partnerships
• Three risks that need confirmation in interviews

Each heading is a bounded task. You can measure how many sources the agent reviewed, how much content it processed, how long it took, and where a human had to correct it.

The agent should summarize source content early rather than carrying full reports into every later step. That one design choice can materially reduce token usage. The source link and the key evidence can be retained, while downstream steps work from a compact research record.

If your firm is considering where research automation fits, our guides can help frame the operating questions. The more important point is to measure the work at task level. “Research” is too broad to cost. “Produce a sourced competitor brief from 12 defined sources” is measurable.

Knowledge debt makes every agent more expensive

Every client project creates intellectual property. Decks, workshop notes, interview transcripts, workplans, deliverables, commercial models, and recommendations accumulate in drives, inboxes, SharePoint sites, and personal folders.

Then someone asks, “Have we done anything similar before?”

The answer is usually a manual search followed by a message to three experienced people.

A Knowledge Agent is designed to read the material the firm produces and answer questions across the corpus. It can identify relevant work, point users to the source document, summarize what was done, and show which partners or teams were involved.

This agent has a major benefit, but it also has a major cost risk. Knowledge bases can be enormous. If every question causes the model to ingest entire decks and transcripts, the workflow will be slow and costly.

A better architecture separates storage from reasoning. The system indexes and classifies documents when they are created. It stores metadata such as sector, service line, client type, project date, sensitivity level, and document type. When a user asks a question, the agent retrieves only the most relevant excerpts and asks the model to answer from those excerpts.

This gives you three controls:

• You limit the source material sent to the model.
• You improve answer quality by retrieving more relevant evidence.
• You can trace which documents supported the answer.

That is the difference between an AI chat tool and an operating asset. You can see how this fits into the wider Omni platform, where agents, workflows, and business controls are designed around the work your team actually does.

Calculate cost per completed task, not cost per user

Most firms begin with a monthly subscription number. That is useful, but it will not tell you if an agent workflow works commercially.

Instead, build a cost model for one completed business task.

For a Proposal Generation Agent, the unit might be one first-draft proposal. For a Research Agent, it might be one sourced company and market brief. For a Knowledge Agent, it might be one answered question with cited internal documents.

Track five numbers for each workflow:

1. Input tokens per task
   This includes instructions, retrieved documents, source summaries, and previous workflow outputs.

2. Output tokens per task
   This includes outlines, drafts, summaries, quality checks, and revisions.

3. Number of model calls
   A workflow with 14 calls is very different from one with four, even if the final output looks similar.

4. Retrieval volume
   Measure how many documents and excerpts the agent reads. Put a sensible ceiling on both.

5. Human time saved or redirected
   Separate avoided work from review work. If a manager still spends two hours repairing a draft, the workflow has not yet earned the right to scale.

Then compare the AI cost to your fully loaded delivery cost. The goal is not to make every task free. The goal is to make the economics visible before the workflow becomes standard practice.

A $4 AI cost to create a solid research brief can be sensible if it replaces three hours of repetitive analyst work and improves sourcing discipline. A $40 AI cost for a draft that needs three hours of rewrites is a different conversation.

You should also model peak demand. Proposal volume is rarely steady. If your firm submits 12 major proposals in a busy month, calculate the cost of the full workflow under that load. Include retries, document processing, and quality reviews. A pilot used twice a week may look cheap because it has not reached the conditions where complexity shows up.

Design agents with token budgets and decision gates

Most cost overruns can be prevented during workflow design. You do not need to turn partners into AI engineers, but you do need to set operating rules before scaling.

Start with a token budget for each task. That budget can be expressed in dollars, number of model calls, maximum documents retrieved, or a combination of all three. When the workflow approaches the limit, it should stop, flag the issue, and request human direction.

For example, a research workflow may be allowed to review 15 external sources and create a 750-word brief. If it cannot answer a key question after that point, it should identify the gap. It should not keep searching indefinitely.

Use smaller models for classification, extraction, and routing where appropriate. Reserve more capable models for work that requires judgment, synthesis, or client-ready language. A workflow does not need its most expensive model to decide whether a document is a case study or a meeting transcript.

Summarize long documents once, then reuse the summary where it is accurate to do so. Keep the original source available for verification. This reduces repeated processing while preserving traceability.

Build clear handoffs. An agent should know when it is expected to proceed, when it must ask a question, and when it should stop. In consulting, client context and commercial judgment are often too important to infer.

This is also where Omni Advisory can be useful. The practical work is mapping the process, deciding what should stay human, defining the source material, and selecting an outcome that is worth automating.

An Omni Audit turns the idea into a costed plan

Most firms don’t need another AI brainstorming session. They need a clear view of where the manual work sits, which agent is commercially viable, and what it will cost to operate.

A 60-minute Omni Audit does that without a long deck or a generic maturity score. We look at the work moving through your firm, from opportunity to delivery to knowledge capture.

You leave with three outputs:

1. A shortlist of high-value workflows that are realistic to automate.
2. A view of the time, leakage, and delivery constraints behind each workflow.
3. A practical first-agent plan, including the data, controls, and token-cost questions to resolve.

If you are already testing proposal, research, or knowledge tools, this is the right time to pressure-test the operating model. See Omni for consulting firms, then Book a 60-min Omni Audit and bring one workflow you are considering.

Use a worksheet before you build

If you want to do some internal preparation first, Deploy Your First Business Agent is a practical worksheet for choosing one task, defining its inputs and outputs, assigning human approval points, and estimating the value at stake. You can also access the direct version here: download the first-agent worksheet.

Use it to document one workflow in plain language. Who starts it? What information does the agent need? What should it produce? What can go wrong? How many times per month does the task occur? Those answers are more useful than a broad AI roadmap.

Scale the workflow, not the token bill

Agentic AI can remove a meaningful amount of repeated work from a consulting firm. It can give analysts a faster starting point, help partners reuse the firm’s best thinking, and reduce the cost of producing proposals and early-stage research.

But the economics will not take care of themselves at scale.

The firms that get this right will not be the ones with the most agents. They will be the ones that treat every agent as an operating process. They will know the unit of work, the source material, the approval gates, the quality standard, the token budget, and the margin impact.

Start with the process where you already know the pain. A 30-hour proposal cycle. Repeated market research. A pile of valuable project IP that nobody can find. Cost the workflow before you automate it, then build the smallest useful agent around it.

For a focused view of that opportunity, review the AI audit for consulting firms and Book my Omni Audit. In 60 minutes, we can identify where agents can reduce real leakage without creating an uncontrolled AI expense.