DeepSeek announced a 75% price cut last week. The headline promised cheaper AI. The reality is more complicated, and consulting firms need to understand what’s actually happening before they spin up new agent workflows.

Lower model prices don’t mean lower bills. They mean higher usage. When a single task becomes 10 or 20 API calls instead of one, your cost per workflow stays flat or climbs. The unit price dropped, but the unit count exploded. This is the 100x problem, and it’s already reshaping how firms should think about AI cost management.

If you’re running a consulting practice between $1M and $25M in revenue, you’re probably looking at AI agents to handle proposal generation, research synthesis, or knowledge management. Those are the right targets. But the economics changed in the last 90 days, and the old mental model of “cheaper model = cheaper bill” doesn’t hold anymore.

Why Agent Systems Multiply API Calls

An AI agent isn’t a single prompt. It’s a loop. The agent reads the task, breaks it into steps, calls the model for each step, evaluates the output, decides whether to loop again, and eventually assembles a final result. A proposal generation agent might make 15 calls to produce one deck. A research agent might make 30 calls to synthesize a brief.

Each call is cheaper now. But you’re making more of them. The total cost per task depends on how many loops the agent runs, how much context it carries forward, and whether the workflow includes retrieval, validation, or revision steps.

Here’s what that looks like in a real consulting workflow. You ask a proposal agent to draft a pitch for a new client. The agent pulls past proposals from your knowledge base (three calls to embed and retrieve). It reads the RFP and extracts requirements (two calls). It drafts an executive summary (one call). It checks the draft against your pricing model and case studies (four calls). It revises based on tone and structure rules (two calls). It formats the output and appends standard terms (one call). That’s 13 calls for one proposal draft.

At the old OpenAI pricing, 13 calls added up fast. At the new DeepSeek pricing, each call is 75% cheaper. But if the workflow was previously designed to minimize calls because they were expensive, you’ll redesign it now to maximize quality because calls are cheap. You’ll add validation loops. You’ll add revision steps. You’ll add retrieval from more sources. The call count climbs to 20 or 25, and your total cost per proposal stays roughly where it was.

This isn’t a bad thing. You’re getting better output. But it’s not the 75% savings the headline promised.

The Consulting Firm Use Case

Consulting firms leak $80K to $300K per year on repeated manual work. The three biggest buckets are proposal time, research synthesis, and knowledge management debt. AI agents can recover most of that leakage, but only if you track the right metrics.

Proposal generation is the clearest example. A senior consultant spends 20 to 40 hours on a major proposal. Half that time is formatting, pulling case studies, and rewriting boilerplate. The other half is strategy and positioning. An agent can’t do the strategy work, but it can handle the mechanical assembly. If you’re writing 12 major proposals a year, that’s 240 to 480 hours of senior time. At a $200 blended rate, you’re spending $48K to $96K on proposal assembly.

A Proposal Generation Agent pulls past proposals, case studies, and pricing into a tailored draft. It doesn’t write the strategy section, but it assembles the rest of the deck in 15 minutes instead of 15 hours. You review it, add the strategic narrative, and ship. The agent made 20 API calls to produce that draft. At current pricing, those 20 calls cost about $0.80. Even if the workflow scales to 30 calls per proposal, you’re spending $1.20 per draft. Over 12 proposals, that’s $14.40 in API costs to recover $48K to $96K in labor.

Research synthesis is similar. Every engagement starts with secondary research. You’re pulling industry reports, competitor analysis, regulatory context, and financial data. A junior consultant spends two weeks on this. The output is a 10-page brief that gets read once and filed. If you’re running six engagements a year, that’s 12 weeks of junior time. At a $120 blended rate, you’re spending $57,600 on research that could be automated.

A Research Agent runs structured industry and company research at the start of every engagement. It queries databases, summarizes sources, and produces a one-page brief with citations. The agent makes 25 to 35 calls per brief. At $0.04 per call, that’s $1.00 to $1.40 per research package. Over six engagements, you’re spending $6 to $8 in API costs to recover $57,600 in labor.

Knowledge management is the hardest problem and the biggest opportunity. Every project produces decks, memos, models, and transcripts. Almost none of it is reusable because no one can find it. A partner asks, “Didn’t we do a pricing model for a SaaS client last year?” The answer is yes, but it’s in someone’s Google Drive, and the person who built it left the firm. You rebuild it from scratch. This happens dozens of times a year across the firm.

A Knowledge Agent reads every document the firm produces and answers questions across the corpus. It’s a retrieval system with a conversational interface. When a partner asks about the SaaS pricing model, the agent finds it, summarizes it, and links to the file. The agent makes 10 to 15 calls per query. At $0.04 per call, that’s $0.40 to $0.60 per question. If the firm runs 200 queries a month, you’re spending $80 to $120 in API costs to eliminate the repeated work that costs the firm tens of thousands in duplicated effort.

These are the workflows where lower model prices should translate to real savings. But only if you’re tracking total workflow cost, not just the unit price per call.

What Changed in the Last 90 Days

The price drop from DeepSeek and the competitive response from OpenAI and Anthropic changed the design constraints for agent systems. When API calls were expensive, you designed workflows to minimize calls. You batched requests. You cached aggressively. You avoided loops. The result was agents that worked, but didn’t work well. They cut corners because every corner was expensive.

Now that calls are cheaper, you can design for quality instead of cost. You can add validation loops. You can retrieve from multiple sources. You can revise outputs until they meet a quality threshold. The agent makes more calls, but the output is better. The total cost per task stays flat or drops slightly, and the output quality jumps.

This is the trade-off consulting firms need to understand. Lower prices don’t mean lower bills. They mean better workflows at similar cost. If you’re building agents now, design for quality. Add the loops. Add the retrieval. Add the validation. The incremental cost is negligible, and the output improvement is substantial.

The other change is that agent usage is climbing faster than anyone expected. When calls were expensive, firms were conservative about deploying agents. They picked one or two high-value workflows and stopped. Now that calls are cheap, firms are deploying agents across more workflows. Proposal generation, research synthesis, knowledge management, meeting prep, client reporting, and internal comms. Each workflow is cheap, but the aggregate usage climbs.

We’re seeing firms that started with one agent six months ago now running five or six. The total API bill is higher than it was, even though the unit price dropped. This isn’t a problem if the agents are recovering more leakage than they cost. But it does mean you need to track usage at the workflow level, not just the model level.

How to Track Workflow Costs

Most firms don’t have a system for this yet. They’re tracking API usage in aggregate, but they’re not breaking it down by workflow or by task. That’s fine when you’re running one or two agents. It doesn’t scale when you’re running six or eight.

You need three metrics. Cost per task, tasks per month, and leakage recovered per task. Cost per task is the total API spend for one completed workflow divided by the number of tasks. For a proposal agent, that’s the total API cost for all proposal drafts divided by the number of proposals. Tasks per month is how often the workflow runs. Leakage recovered per task is the labor cost you avoided by using the agent instead of doing the work manually.

If your proposal agent costs $1.20 per draft, runs 12 times a year, and recovers $4,000 to $8,000 in labor per proposal, the ROI is obvious. If your research agent costs $1.40 per brief, runs six times a year, and recovers $9,600 in labor per engagement, the ROI is even clearer. If your knowledge agent costs $0.60 per query, runs 200 times a month, and eliminates duplicated work worth $120 per query, you’re recovering $24,000 a month in leakage for $120 in API costs.

These are the numbers that matter. Not the unit price per API call. Not the aggregate monthly bill. The cost per workflow, the frequency of the workflow, and the leakage recovered.

If you want a structured way to map this for your firm, we built a worksheet that walks through the exercise. Deploy Your First Business Agent is a 12-page guide that helps you pick the right workflow, estimate the cost per task, and calculate the ROI. It’s free, and it’s designed for consulting firms that want to move from theory to deployment in the next 30 days.

The Omni Audit for Consulting Firms

If you’re not sure which workflows to prioritize, or if you want someone to walk through the cost model with you, book a 60-min Omni Audit. It’s a working session, not a sales call. You’ll walk away with three outputs: a leakage map that shows where your firm is losing time and money, a prioritized list of agent workflows ranked by ROI, and a 90-day deployment plan with cost estimates and milestones.

We run these audits for consulting firms between $1M and $25M in revenue. The session is 60 minutes. We don’t use a deck. We ask questions, map your workflows, and build the plan in real time. If you want to move forward, we can start building the first agent the following week. If you don’t, you keep the plan and the leakage map.

The audit is designed for consulting firms that know they need AI agents but don’t know where to start. It’s also designed for firms that already deployed one or two agents and want to scale to five or six without multiplying their API bill by five.

What This Means for the Next 12 Months

Lower model prices are permanent. Agent usage will keep climbing. The firms that win are the ones that track workflow costs and design for quality instead of call minimization. The firms that lose are the ones that assume cheaper models mean cheaper bills and don’t instrument their usage.

If you’re running a consulting practice, you’re already thinking about agents. You’ve probably tested a few workflows. You’ve probably seen the ROI on proposal generation or research synthesis. The next step is to scale from two agents to six, and to do that without losing control of your costs.

That requires a system. You need to track cost per task, tasks per month, and leakage recovered. You need to instrument your workflows so you know which agents are paying for themselves and which ones are running up bills without recovering enough labor. You need to design new workflows with quality in mind, not just cost minimization.

This is the work we do with consulting firms through Omni Ops. We build the agents, instrument the workflows, and track the metrics. We don’t hand you a chatbot and walk away. We build the system, deploy it into your workflows, and iterate until the ROI is clear.

If you want to see what that looks like for your firm, book your Omni Audit and we’ll map it out in 60 minutes. You’ll leave with a plan, a cost model, and a clear next step. No deck, no pitch, no follow-up emails. Just the plan.

Lower prices are driving higher usage. The firms that understand this will recover $80K to $300K in leakage over the next 12 months. The firms that don’t will wonder why their API bill went up even though the unit price went down. The difference is instrumentation, and the time to build that system is now.