The economics of replacing roles vs augmenting them

"What's the real number if I move from 8 SDRs to 3 SDRs plus an agent? Including the hidden costs nobody puts in the vendor deck?"

That's the question I get most from CFOs and operators trying to model their next-year plan. Not "should I use AI." The structural question: when is replacement the right move, when is augmentation, and what is the TCO honestly?

This is the framework. With the formula. And three worked examples.

TL;DR

The 4 factors

Before you build a model, run any candidate role through these 4 questions.

1. Error tolerance. What's the cost of a single wrong output? If a wrong response to a customer service ticket costs $30 in customer goodwill and 4 minutes of escalation cleanup, error tolerance is high. If a wrong AR reconciliation entry causes a $40K audit finding 6 months later, error tolerance is zero. High-tolerance roles are replace candidates. Zero-tolerance roles are augment-with-strong-verification, not replace.

2. Structured input. Does the work start from a well-defined input? AR reconciliation starts from an invoice and a bank statement. That is structured. Strategic account renewal starts from a 9-month conversation history, an industry context, and a relationship. That is unstructured. Structured input is a replace candidate. Unstructured input is an augment candidate.

3. Output verifiability. Can you tell at a glance whether the output is correct? A line item that reconciles is verifiable. A reply to a sales prospect that says "yes, our product handles this" is verifiable. A piece of strategic advice given to a customer is not verifiable until 8 months later when the renewal comes around. Verifiable = replace candidate. Hard-to-verify = augment.

4. Customer-facing surface. Does the output go directly to a customer or end-user? Customer-facing replacement is doable but expensive. You need the escalation layer, the brand-voice match, the satisfaction monitoring. Internal-facing replacement is much cheaper. If you're choosing where to start, start with the internal-facing roles. Replace customer-facing only after the internal motion is solid.

A role that scores green on all 4 is a strong replace candidate. A role that scores green on 2 to 3 is augment. A role that scores red on 3 or 4 is "leave the human alone for now."

The honest TCO formula

Here is the 7-line TCO model you should be using. The vendor deck shows you line 1 and line 2. The other 5 are the ones that make the math actually balance.

Line 1. Direct API / vendor cost. The price of the AI tokens or the per-seat license. Easy to model. For most agent workflows this is $40 to $400/month per "replaced FTE equivalent."

Line 2. Salary saved (if replace). The fully-loaded comp of the role being replaced. Include benefits, taxes, equipment, training.

Line 3. Build / integration cost. Wiring the agent into your existing systems. CRM integration, calendar integration, the orchestrator layer, the verification queue. Usually $8K to $40K up-front per workflow. Often $200 to $1,500/month ongoing maintenance.

Line 4. Verification labor. A human checks a sample of the agent's outputs every week. This is a non-zero cost. For a customer-facing agent, this is 4 to 8 hours per week of a competent operator's time. For an internal-facing agent it's 1 to 2 hours per week.

Line 5. Error / escalation cost. When the agent gets it wrong, somebody has to fix it. Quantify the error rate from your pilot. Multiply by the cost of fixing each error. For a support agent, that's the cost of the second-touch ticket plus the satisfaction hit. For a reconciliation agent, it's the cost of the audit finding. This is the line item every "AI saves $1.2M" pitch leaves out.

Line 6. Reputation / brand risk reserve. A budget line for the case where the agent does something embarrassing publicly. Usually 2 to 5% of the saved salary, set aside as reserve. Most teams skip this and regret it once.

Line 7. Re-skilling / opportunity cost on the human side. If you redeploy the people who used to do this work, you have a training cost. If you reduce headcount, you have a severance cost and a culture cost. Either way it's real.

The honest TCO for replacement is Line 1 + Line 3 + Line 4 + Line 5 + Line 6 + Line 7, minus Line 2.

For augmentation, you have Line 1 + Line 3 + Line 4 + a smaller Line 5, and no Line 2 savings.

Now you can compare them apples-to-apples.

Worked example 1: outbound SDR team

The scenario: 8-person outbound SDR team. Each rep is fully loaded at $85K/year (salary + benefits + tools + management overhead). Total annual cost: $680K. They generate 2,400 qualified meetings a year.

The proposal: replace 5 of them with an outreach agent stack. Keep 3 senior SDRs for handoffs and complex accounts.

Replace TCO (annual):

Net annual savings: $425K minus ($24K + $16K + $15K + $3.6K + $12K + $30K) = $324.4K/year.

Pipeline impact: if the agent stack generates 90% of the meetings per saved SDR (best case), pipeline drops by 10% × 5 = ~150 meetings/year. At an average $4K influenced revenue per meeting, that's a $600K influenced revenue gap. Net dollar impact is now negative.

Better proposal: keep all 8 SDRs, give each one an agent that triples their output. Same payroll. 2x to 3x the qualified meetings. Augmentation wins on this role for most companies.

The lesson: when the role generates revenue, augmentation almost always beats replacement, because you're not capacity-constrained on the salary, you're capacity-constrained on the output.

Worked example 2: support tier 1

The scenario: 6-person tier 1 support team handling 12,000 tickets/month. Each rep fully loaded at $65K/year. Total $390K/year.

The proposal: replace 4 of them with a support agent that handles tickets with deterministic, documented answers. Keep 2 humans for escalations and edge cases.

Replace TCO (annual):

Net annual savings: $260K minus ($18K + $23K + $24K + $126K + $7.8K + $20K) = $41.2K/year.

That's a much smaller number than the vendor deck would have shown you. The deck would have led with "$260K saved." The honest number is $41.2K, and it could go negative if the error rate creeps from 5% to 8%.

Better proposal: use the agent to handle the 40% of tickets that are deterministic, route the rest to humans, keep all 6 reps. You get the cycle-time win, the CSAT score, and the math works out positive without the error-cost cliff.

Worked example 3: AR reconciliation

The scenario: 2-person AR team. Each fully loaded at $55K/year. Total $110K/year. They reconcile 800 invoices/month.

The proposal: replace one of them with an agent that does the reconciliation work. Keep one human for exceptions and audit-trail review.

Replace TCO (annual):

Net annual savings: $55K minus ($6K + $9K + $8K + $10K + $1K + $15K) = $6K/year.

After 2 years, the redeployment line drops off and savings hit $21K/year. That's a real positive ROI, on a role that scored all green on the 4-factor test.

The lesson: AR reconciliation is a textbook replace candidate (structured input, verifiable output, internal-facing, moderate error tolerance with a reserve). And the honest TCO is still single-digit thousands of dollars in year one, not the $55K the vendor deck would suggest.

The pattern

Three worked examples. Three different answers.

The decision is per-role. Not per-company. Not per-"AI strategy." A real CFO model has each role scored on the 4 factors, the 7-line TCO done honestly, and a per-role recommendation.

For how this fits the broader org-design picture, see the cornerstone post.

What changes about hiring

If your model says "replace 3 SDRs," that does not mean "fire 3 people." It means "hire 3 fewer next year." The framing matters operationally and culturally.

Companies that ran the "fire and replace" playbook in 2023 and 2024 mostly ended up rehiring within 18 months because the work didn't actually go away. Companies that ran the "freeze and redeploy" playbook kept the institutional knowledge and absorbed the agent stack alongside the team. The second group is doing better.

The advice: if your TCO model produces a "replace" recommendation, time it to natural attrition. Don't force the cut. Use the saved capacity to take on more work, not to ship the same work with fewer people.

The roles to never replace

There are 4 categories of work an agent should never own end-to-end, regardless of what the TCO model says.

The customer's first conversation with you. Sales, intake, onboarding kickoffs. The relationship is the asset.

Final hiring decisions and performance reviews. Accountability has to live with a specific human.

Anything legal or regulated where the wrong answer compounds. Tax filings, compliance reports, legal advice.

Anything where a single wrong output ends a customer relationship. Refund decisions over a threshold, account closures, public statements.

These look like cost centers. They're actually the brand. Replace them and the cost shows up in the next quarter's churn number, not the headcount line.

What to do in week one

You don't need a 90-day model. Here's the path.

Day 1. Pick the role you're most tempted to "AI away." Run it through the 4-factor test.

Days 2 to 3. Build the honest 7-line TCO model for that role. Resist the urge to skip lines 4 through 7. Get a real number for line 5 (the error cost) by talking to whoever currently handles errors.

Day 4. Compare against augmentation: what if you kept the headcount and gave each person an agent? What's the output gain?

Day 5. Pick the better option. If augmentation wins, build the agent layer and keep the team. If replacement wins on a role that scored 4/4 green, plan it as freeze-and-redeploy, not fire-and-replace.

Weeks 2 to 4. Build the agent. Measure error rate. Recalibrate the TCO with real data after 30 days.

Month 3. Have the board conversation. With a model that has all 7 lines in it. Your numbers will be smaller than the vendor decks promised. They will also be defensible, which the vendor numbers are not.

Next up

Next post is the conversation nobody is having and everyone is about to need: compensation and equity for hybrid human-AI teams. What happens to comp bands when an agent does 40% of the role? Who gets the equity? Who gets the performance bonus? After that we'll cover vendor selection, build vs buy vs orchestrate, for agentic AI.

If you want this TCO model run for your business, see the blueprint catalog or email christine@operatoriq.io. Email only, no calls.

Cheers, Christine