AI tools return an average of $3.50 per $1 invested, per a 2024 IBM study of 3,000+ enterprise deployments. Most jobs aren't being eliminated — specific tasks within jobs are. The roles quietly disappearing are the ones built almost entirely around moving information from one box to another. If that describes your job, this article is for you. (Sorry.)
Every few months, a new study drops confirming that AI will automate somewhere between "some jobs" and "all of civilization." Goldman Sachs says 300 million roles are exposed. The World Economic Forum says 85 million jobs will be displaced by 2025 — a number that has since been quietly revised, because predicting AI's workforce impact is apparently harder than it looks. Meanwhile, every CEO on LinkedIn is posting about their company's "exciting AI transformation journey" while their HR team processes layoff paperwork in another tab.
Here's what's actually happening — with numbers, not vibes.
The ROI Formula — Because Vendors Won't Show You This
AI vendors love to quote ROI. They just hate including the full cost side of the equation. Here's the honest version — napkin-ready, no consulting fee required.
Let's Run It — 10-Person Team
- Task automated: document drafting, research, email triage (the stuff everyone hates doing anyway)
- Time AI handles: 2 hours/person/day
- Fully-loaded cost per employee: $75,000/yr ≈ $36/hr
- Annual time savings: 10 × 2 hrs × 250 days × $36 = $180,000/yr
- GPT-4o API cost at that volume: ~$9,000/yr
- Setup + integration (one-time): $12,000
| Year | Gross Savings | Total Cost | ROI |
|---|---|---|---|
| Year 1 | $180,000 | $21,000 | 757% |
| Year 2+ | $180,000 | $9,000 | 1,900% |
1,900% ROI sounds made up. It isn't. The economics of replacing human labor hours with API calls are genuinely that lopsided — when aimed at the right tasks. That caveat is doing a lot of heavy lifting in this sentence, which is why there's an entire section on hidden costs below.
Where AI Prints Money (and Where It Just Prints Apology Emails)
Not all use cases are equal. Some deliver 800% ROI in year one. Others deliver 40% ROI and a customer satisfaction score that looks like a stock chart in 2022.
| Use Case | Avg. Time Saved | Annual Tool Cost | 1-Year ROI (10 staff) | Job Risk |
|---|---|---|---|---|
| Email drafting & summarization | 1.5 hrs/day | $3,000–8,000 | 400–700% | Low — augments |
| Tier-1 customer support | 3 hrs/day | $12,000–30,000 | 200–400% | High — replaces |
| Code generation & review | 2 hrs/day | $15,000–40,000 | 150–300% | Medium — augments |
| Data analysis & reporting | 2.5 hrs/day | $5,000–15,000 | 500–900% | Medium — augments |
| Contract first-pass review ⭐ | 3.5 hrs/day | $8,000–25,000 | 600–1,000% | High — replaces |
| Content moderation | 4 hrs/day | $20,000–60,000 | 300–600% | High — replaces |
The pattern is obvious once you see it: AI crushes roles where a human spends most of their day taking input → applying a rule → producing output. It struggles badly anywhere the job requires reading a room, navigating ambiguity, or being held personally accountable when things go sideways. *(Which, come to think of it, is exactly the work most people find meaningful. Funny how that works.)*
A 2024 MIT study found that workers using AI tools completed tasks 26% faster and produced higher-quality output than non-AI counterparts. That gap widened every single quarter. The uncomfortable implication: two people with the same job title are now doing meaningfully different amounts of work depending on whether they've figured out AI. Managers haven't quite caught up to this yet. They will.
Which Jobs Are Actually Getting Replaced (Be Honest)
Every article about AI and jobs says "don't worry, it's just changing roles, not eliminating them." Some of those articles are written by people whose jobs are not changing. Here's what's actually happening at companies that have already deployed at scale.
Already Gone (or Going Fast)
- Tier-1 customer support: Klarna's AI assistant now handles the equivalent of 700 full-time agents — a fact Klarna announced proudly in a press release, which is a bold move. 60–70% of tier-1 queries at major platforms resolve without a human ever reading them.
- Basic data entry: McKinsey's 2023 automation index pegged 87% of data-entry tasks as fully automatable right now. Not "eventually." Now. The only reason many of these jobs still exist is that companies haven't gotten around to the migration yet.
- First-pass financial reporting: Goldman Sachs cut analyst time on first-draft report generation by 80% after deploying LLM tooling. The analysts didn't disappear — they just stopped doing the part of the job that took 80% of their time.
- Contract review (first pass): Law firms Allen & Overy and Linklaters both reported 85%+ reductions in time spent on initial contract review. Junior associates are now doing in two hours what used to take two days. [Checks notes on what junior associates were billing at.] Hmm.
Genuinely Safe (For Now)
- Skilled trades — AI cannot rewire your house, and it's not trying to
- Roles built on long-term trust relationships — therapists, advisors, account managers who actually know their clients
- Creative direction — deciding what to make and why, not executing it
- High-accountability judgment calls — anything where a human signature still matters legally or ethically
IBM surveyed 3,000 executives in 2023 about their workforce plans. The finding everyone cited: 40% of workers would need reskilling within 3 years. The part fewer people mentioned: "reskilling" is corporate for "your current job description is about to change substantially, and we hope you're okay with that." Not fired. Just... different. Whether that's reassuring depends entirely on how much you liked your current job description.
The Hidden Costs That Make Your CFO Cry
Here's where the 1,900% ROI number gets a reality check. The vendor demo always shows the savings. It never shows the slide about what it costs to actually get there.
| Hidden Cost | Typical Range | Usually Budgeted? |
|---|---|---|
| Integration development | $8,000–40,000 | Sometimes |
| Prompt engineering & ongoing maintenance | $15,000–40,000/yr | Rarely |
| Accuracy monitoring & QA systems | $5,000–20,000/yr | Rarely |
| Employee retraining | $4,500–8,000/person | Almost never |
| Total hidden costs (10-person team, yr 1) | $48,000–148,000 | Almost never |
Prompt engineering is not set-and-forget. LLM outputs degrade silently when the underlying model updates — and models update constantly. Someone has to own this. That person costs money. Integration takes 3–6 months, not the "2–4 weeks" in the sales deck. And McKinsey puts the median retraining cost at $4,500–$8,000 per employee — a line item that appears in approximately zero AI business cases written by the people buying the software.
Run the ROI formula with these numbers included. You'll still usually end up positive — often dramatically positive. You'll just do it with your eyes open instead of your fingers crossed.
Frequently Asked Questions
How long does it take to actually see ROI from an AI rollout?
For simple use cases like drafting and summarization, 60–90 days is realistic. Customer support automation takes 4–6 months once you factor in model training, QA cycles, and the inevitable edge cases your bot handles with spectacular wrongness. Complex deployments — legal, financial compliance — run 9–18 months before ROI turns positive. Anyone promising faster than that is selling you the demo, not the deployment.
What does it actually cost to deploy an LLM at a small business?
For a 5–20 person company using existing APIs (OpenAI, Anthropic, Google), expect $500–3,000/month in model costs depending on volume, plus $5,000–20,000 one-time for custom integrations. Off-the-shelf tools — Cursor, Notion AI, Intercom's Fin — run $50–500/month with almost no setup. Start there. Custom builds are for when you've proven the use case works, not before.
Which jobs are actually most at risk right now?
Any role where 70%+ of the day is spent processing structured information into predictable outputs: data entry, tier-1 support, first-pass document review, basic financial reporting. The World Economic Forum's 2025 Future of Jobs report named clerical and administrative support as the single largest category of expected decline through 2030. Not "at risk of disruption" — declining. Present tense.
Is AI replacing workers or just changing their jobs?
Both, depending on the role — and companies are not always being straight with you about which one is happening. For knowledge workers doing creative or relationship-heavy work, AI adds capability without cutting headcount. For high-volume procedural roles, it's replacing headcount — documented in actual hiring freezes at Salesforce, Google, and Klarna, not just theoretical models.
How do I calculate AI ROI for my specific situation?
Start with the task that costs you the most hours per week and produces the most predictable output. Multiply hours × people × hourly rate × 250 workdays. That's your savings ceiling. Then price a tool that targets exactly that task. If the ratio isn't at least 5:1, pick a different task. Our AI ROI calculator runs this math with your actual numbers.
What mistakes do companies always make when deploying AI?
Three, every time: deploying on tasks where the quality bar is too high for current models (and finding out the hard way), omitting hidden costs from the ROI projection (integration, prompt maintenance, retraining), and not assigning one internal person to own the deployment obsessively. The companies hitting 500%+ ROI have that person. The companies hitting 40% and blaming the technology usually don't.
Does firing someone and replacing them with AI actually save money after severance?
For high-volume, low-specialization roles like tier-1 support — yes, the math usually works even with severance. For specialized knowledge workers above $80,000/year, the risk-adjusted math often favors augmentation over replacement. A failed AI deployment, degraded output quality, customer complaints, and eventual rehiring can easily cost more than you saved. Most CFOs running honest models reach the same conclusion: replace the task, keep the person.
Here's the exercise worth doing this week: find the task your team does on autopilot — the one everyone vaguely resents but no one has bothered to fix. Run the math on it. Hours × people × rate × 250. Then spend one afternoon testing a tool against it. The ROI calculation either justifies itself immediately or it doesn't. Either answer is useful information.