Agentic AI doesn’t automatically get cheaper at scale

There is a tempting story in consulting right now. Build an AI agent for research, proposal work, or delivery support. Give it more engagements. Spread the setup cost across the firm. Watch the unit cost fall.

That story is incomplete.

A recent Gartner view reported by Computer Weekly makes an important point. Agentic AI doesn’t necessarily benefit from the usual economies of scale. More agent activity can mean more model calls, more tokens, more tool usage, more workflow exceptions, and more human review.

For a consulting or advisory firm doing $1 million to $25 million in annual revenue, this matters early. You don’t need thousands of users to create a cost problem. A 20-person firm can run an expensive AI delivery workflow if every proposal launches multiple research tasks, every task uses a premium model, and a senior consultant still has to inspect every output line by line.

The issue isn’t that agents are a bad investment. The issue is treating them like low-cost automation before you understand their operating model.

A well-designed agent can reduce repeated work and make your firm’s knowledge more useful. A poorly designed agent adds a new technology bill while moving the same work into review queues. That is not leverage. It’s another layer of overhead.

The practical answer is to price and pilot each workflow with guardrails from day one:

• A defined token and model budget per task
• Clear routing rules for cheap versus premium models
• Limits on retries and research depth
• A human-review standard matched to the risk of the work
• A way to measure cost per proposal, engagement, and reusable asset

That discipline is especially important in consulting because the outputs are client-facing. A wrong fact in a generic internal summary is inconvenient. A wrong fact in a strategy deck, commercial proposal, or board recommendation can cost trust.

The expensive work agents are being asked to fix

Most consulting firms don’t have a productivity problem in the abstract. They have repeated pockets of high-value people doing low-leverage work.

Three are common.

Proposal development absorbs senior time

A major proposal often takes 20 to 40 hours. The visible task is writing slides and shaping a narrative. The hidden work is more expensive.

A partner searches for relevant case studies. A manager finds old pricing tables in a folder. Someone copies credentials from an old deck, then asks whether those credentials are current. A senior person rewrites the executive summary because the first draft sounds generic. The team tries to match the prospect’s language while protecting margin in the proposed scope.

If the firm wins enough work, leaders can tolerate this. But win rate can be acceptable while cost of sale is brutal.

The Proposal Generation Agent (Omni ops) is designed for this workflow. It can pull approved past proposals, case studies, capability statements, pricing ranges, and client context into a structured first draft. It doesn’t replace the partner’s commercial judgment. It gives the partner a stronger starting point and shows the source material behind the draft.

The trap is assuming every proposal should receive unlimited AI effort. It shouldn’t.

A $30,000 discovery proposal and a $500,000 transformation pitch don’t deserve the same model budget, review steps, or research depth. Your workflow needs a route for each.

Research gets repeated engagement after engagement

At the beginning of an engagement, teams often spend days gathering secondary research. They map competitors, read annual reports, scan industry reports, collect public customer signals, and assemble a point of view.

Some of that research is necessary. Much of it is repeated.

A retail consulting team may rebuild a market overview for every new client. A workforce advisory firm may re-examine the same regulatory changes for several clients in one quarter. A technology consultancy may reassemble standard vendor comparisons because nobody can find the last good version.

The Research Agent (Omni ops) runs a structured version of this early-stage work. It gathers sources, produces summaries, identifies open questions, and creates a one-page brief for the engagement team. The useful version of this agent does not simply scrape the web and generate confident prose. It retains source links, separates facts from interpretation, and flags where human judgment is still required.

Yet research agents can become one of the most expensive patterns in an AI stack. Open-ended research prompts invite long retrieval loops and repeated calls to premium models. If each consultant can ask for unlimited deep research, costs can climb without any relationship to project value.

Your knowledge stays trapped inside delivery files

Every engagement creates intellectual property. There are workshop transcripts, diagnostic findings, benchmarks, recommendations, operating models, plans, and client-specific lessons.

Then the work lands in SharePoint, Google Drive, a project folder, or the laptop of the person who built it.

The firm pays for the same insight twice when the next team cannot find it. It pays again when an experienced manager leaves and their practical knowledge goes with them. It pays again when a new proposal starts from scratch because relevant examples are buried in old delivery material.

The Knowledge Agent (Omni ops) reads approved decks, documents, and meeting transcripts, then answers questions across the firm’s corpus. A consultant can ask, “What operating model patterns have we used for mid-market manufacturers?” and get an answer with references to the underlying materials.

This is a useful capability, but it needs access controls, source permissions, retention rules, and answer-quality checks. A knowledge agent that can see everything is not automatically a good idea. Client confidentiality and commercial sensitivity have to shape the design.

For firms that want to map these opportunities against their current systems, See Omni for consulting firms. The starting point is not buying an agent. It is deciding where the work, knowledge, risk, and cost sit now.

Why scale can raise the bill

Traditional software has a familiar cost profile. You pay for implementation, licensing, support, and perhaps incremental users. As usage rises, the cost per transaction often falls.

Agent workflows are different because each completed piece of work may trigger variable consumption.

Consider a proposal workflow. One brief might involve:

1. Reading the opportunity notes and CRM context.
2. Retrieving relevant case studies and credentials.
3. Searching for industry facts and prospect signals.
4. Drafting a proposal structure.
5. Drafting individual sections.
6. Checking claims against source material.
7. Revising after partner feedback.
8. Producing a final document and presentation outline.

Each step can use tokens, models, retrieval systems, external tools, and workflow orchestration. If the agent retries steps, follows weak search results, or runs several premium-model reviews, the cost rises quickly.

Then add human review. A partner may save time on a first draft but still spend 90 minutes correcting tone, validating case-study claims, and changing the commercial approach. That might be entirely worthwhile for a strategic pitch. It is less worthwhile for routine tenders if the process doesn’t reduce review effort over time.

This is why cost needs to be measured as more than an AI subscription. Look at:

• Cost per completed proposal draft
• Cost per research brief accepted by the delivery lead
• Cost per usable answer from the knowledge base
• Senior review minutes per workflow
• Rework rate after human review
• Gross margin effect on the work the agent supports

The numbers will vary by firm, model choice, and scope. In advisory work, the bigger risk is often not a shocking model invoice. It is a workflow that quietly consumes partner attention while creating the appearance of automation.

Across firms of this size, we usually see annual leakage from repeated proposal effort, repeated research, and inaccessible knowledge in the $80,000 to $300,000 range. That is not all recoverable through AI. Some work requires expert judgment. But it is enough to justify disciplined workflow design.

Put guardrails into the workflow, not in a policy document

Most firms start with a broad AI policy. They tell staff to avoid confidential data, check outputs, and use approved tools. That is necessary, but it doesn’t manage operating cost.

Cost guardrails need to exist inside the workflow.

Set a budget by job type

Start with the value of the job, not the excitement around the technology.

A low-value qualification brief might have a small fixed budget and use a fast, lower-cost model. A large strategic proposal may justify more research, a stronger model, and a partner-led review. A recurring market scan should have a firm ceiling on sources, iterations, and output length.

The agent should know when it has reached the limit. It can then return a useful partial result, identify what is missing, and ask for approval before spending more.

That is far better than an agent autonomously running another six research loops because it thinks the answer could improve.

Route work to the right model

Not every step needs the most capable model available.

Extracting names, dates, document titles, and standard credentials may be handled by a lower-cost model or a deterministic process. Comparing complex commercial options, synthesising ambiguous research, or drafting a high-stakes executive narrative may need a stronger model.

Model routing is one of the clearest ways to control cost without lowering quality. The standard should be simple. Use the least expensive method that can reliably complete the step.

This is part of how we think about Omni ops. The goal is not to put AI in every process. It is to build operational workflows where the handoffs, controls, and accountability are visible.

Cap retries and research expansion

Agents fail differently from people. A consultant might recognise after ten minutes that a public source is weak. An agent can spend through dozens of calls attempting to resolve the same uncertainty.

Set retry limits. Set a maximum number of sources for standard research. Define which source types count as credible for a given engagement. Require an escalation when the agent cannot validate a critical fact.

Research depth should be selected, not assumed. A one-page market brief for an introductory call may need six credible sources. A board-level market assessment may need a different workflow, a larger budget, and named human responsibility for verification.

Make review proportional to risk

A human in the loop is not a guardrail if their role is undefined.

For each workflow, decide who reviews what and why. A junior consultant may check formatting and source completeness. A manager may validate synthesis and delivery relevance. A partner may approve positioning, pricing logic, and claims that affect the firm’s reputation.

The review should focus on the decisions the agent can’t own. If senior people are still rewriting every paragraph, the workflow isn’t ready to scale. That is useful feedback. It may mean the inputs are weak, the knowledge base is poorly structured, or the output is trying to solve too much at once.

If you want a practical worksheet before investing in a build, Deploy Your First Business Agent is designed to help you select one workflow, define the operating rules, and identify where human approval belongs. You can also access the direct worksheet download for your planning session.

What a guarded proposal agent looks like

Take a common use case, responding to a new advisory opportunity.

The partner adds the opportunity brief, known stakeholders, target scope, and expected fee range. The Proposal Generation Agent classifies the opportunity by value and complexity. That classification sets the workflow budget.

For a standard mid-market proposal, the agent may be allowed to retrieve only approved internal material. It searches the firm’s case-study library, service descriptions, pricing patterns, and relevant prior proposals. It drafts an outline, recommended team, assumptions, indicative workplan, and a first executive summary.

It tags every reusable claim with its internal source. It does not invent client outcomes, industry benchmarks, or specialist credentials. If it cannot find a relevant proof point, it marks the gap.

The manager reviews the draft against a short checklist:

• Is the problem definition accurate?
• Are all case-study claims approved and relevant?
• Does the workplan fit the opportunity?
• Are pricing assumptions commercially sound?
• Does the proposal contain any unsupported statement?

For a larger or more complex opportunity, the workflow might permit external research. It may run a limited company and market scan, then add a short evidence section for the partner. It should not automatically turn that research into unverified claims in the proposal.

The output improves with use because the agent is connected to an increasingly structured corpus of approved materials. But the cost only improves if you keep the workflow bounded. More proposals do not guarantee lower unit cost. Better retrieval, model routing, reusable templates, and less senior rework do.

That distinction is the whole point.

Pilot one workflow before building a firm-wide agent layer

Don’t start by promising an AI assistant for every consultant. Start with one repeatable workflow that has a clear baseline.

Proposal generation is often a good first pilot because the firm can measure hours, outputs, review time, and win-related quality. Structured research can also work well where the same market or client types recur. Knowledge management is valuable, but it often requires more preparation around permissions, document quality, and taxonomy.

For a 30-day pilot, set targets that can be checked:

• Use the workflow on 5 to 10 real opportunities or engagement starts.
• Record AI consumption and tool costs for each run.
• Track preparation hours before and after.
• Record manager and partner review minutes.
• Score source quality, factual accuracy, and usefulness.
• Set a maximum cost per completed output.
• Document exceptions rather than hiding them.

You are not trying to prove that AI can generate text. That has already been demonstrated. You are testing whether a controlled workflow can reduce repeated work while protecting client quality and margin.

The broader Omni platform can support the move from a single pilot to connected operational agents, but only after the firm understands its workflow economics. Build the control model first. Then scale what works.

Find the leakage before you scale the agent

The firms that get real value from agent workflows aren’t usually the ones with the biggest list of AI tools. They are the ones that know where their people lose time, where their knowledge breaks down, and where an automated workflow requires human judgment.

A 60-minute Omni Audit is built to make that clear. There is no long deck and no generic maturity score. We work through your current workflow and leave you with three practical outputs:

1. The highest-value AI workflow opportunity for your firm.
2. The people, process, data, and system constraints that affect it.
3. A prioritised path to pilot it with cost and review guardrails.

If proposal work, repeated research, or inaccessible delivery IP is costing your firm more than it should, Book a 60-min Omni Audit.

You can also review the AI audit for consulting firms to see the specific areas we assess. The aim is straightforward. Make your next agent useful, measurable, and safe to scale.

Agentic AI can create leverage in a consulting firm. It will not create that leverage just because you run it more often. Price the work. Limit the workflow. Keep people accountable for the judgments that matter. Then expand from evidence, not optimism.

When you’re ready to identify the first workflow worth piloting, Book my Omni Audit.