In short: To implement AI in business without wasting money, start with one costly workflow and a measured baseline, not a company-wide strategy or a tool. Run a time-limited pilot with the people who do the work, redesign the process around it, and set rules for data and review. Scale only when the numbers beat the baseline after all costs.
Plenty of firms are spending without much to show. MIT’s NANDA project reported in 2025 that 95% of organizations it studied saw no measurable business return from generative AI initiatives. Of enterprise-grade systems evaluated, 60% of organizations looked at them, 20% reached the pilot stage and only 5% reached production. Treat that number with caution, since it’s based on 52 interviews, 153 conference survey responses and a review of over 300 public initiatives, the authors themselves call their figures directionally accurate, and the group behind it develops agent infrastructure. But the direction matches what many executives report: experiments are easy, and results are hard.
Step 1: Choose a workflow, not a technology
“We need an AI strategy” tends to generate slide decks. “Our invoice exceptions take 400 staff hours a month” generates a project.
A good first candidate is:
- Frequent, so savings add up and you get data quickly
- Costly, in staff time, errors or delay
- Text- or data-heavy, with material that AI can actually use
- Easy to check, so mistakes are caught
- Low in legal risk, at least to begin with
- Owned by someone who wants the change
Rank candidates using a simple score:
| Workflow | Annual cost | Data ready? | Risk if wrong | Easy to check? | Priority |
|---|---|---|---|---|---|
| Support ticket drafting | High | Yes | Medium | Yes | High |
| Contract review summaries | Medium | Partly | High | Partly | Medium |
| Automated loan decisions | High | Yes | Very high | No | Low for now |
| Internal knowledge search | Medium | No, documents are messy | Low | Yes | After cleanup |
McKinsey’s 2025 survey suggests that most firms aim for efficiency, while high performers also set growth and innovation goals. That’s a useful second stage, but your first project should have a number you can verify quickly.
Step 2: Measure the baseline
Without a baseline, you can’t tell whether anything changed, and people’s impressions are unreliable. In METR’s small randomized trial, experienced developers took 19% longer with AI tools, yet believed they had worked 20% faster.
Record the current state for at least two to four weeks:
- Time per task, including rework
- Volume handled
- Error rate or quality score
- Cost per unit
- Customer or employee satisfaction, where relevant
Keep a comparison group during the pilot if you can. In Brynjolfsson, Li and Raymond’s study of about 5,000 support agents, the gradual rollout made it possible to compare agents with and without the tool. That’s how the researchers could show a 13.8% rise in issues resolved per hour, with no significant change in customer satisfaction, instead of merely asserting it.
Step 3: Decide whether to buy, configure or build
| Option | Fits when | Watch out for |
|---|---|---|
| Use features in existing software | A common need, and data already sits in the system | Limited flexibility; added license costs |
| Buy a specialist product | A well-defined problem others share | Vendor lock-in, demos that don’t reflect your data |
| Configure a general assistant | Text tasks, with human review | Inconsistent quality and governance |
| Build a custom system | Your data or process is a real advantage | Long timelines, maintenance, talent |
Begin with the simplest option that could work. Custom builds look attractive, but in NANDA’s review only 5% of custom enterprise tools reached production. Ask vendors for references in your industry, a pilot on your data and exit terms, including how you’d get your data out.
Step 4: Run a limited pilot
- Scope. One team, one workflow, 6 to 12 weeks.
- Team. Include people who do the work, plus someone from IT, and legal or compliance where needed. Mid-market firms in the NANDA research moved from pilot to implementation in about 90 days, against nine months or more at large enterprises, partly because fewer layers were involved.
- Success criteria set in advance. For example: “20% less handling time with no rise in errors, net of all costs.”
- Human review. Keep a person approving outputs until you’ve seen enough evidence.
- A log of failures. Record where it goes wrong. This is your most useful data.
Step 5: Redesign the work
Dropping AI into an unchanged process usually disappoints. McKinsey’s survey found high performers nearly three times as likely to have fundamentally redesigned workflows, and NANDA identified brittle workflows and poor alignment with daily operations as common failure points.
Ask what should change if drafting takes seconds. Who reviews it? What gets removed? Which approvals still matter? If AI speeds up one step and the next step is a person with a growing queue, the business hasn’t gained anything.
Decide on saved time as well. BCG’s 2026 survey found that 66% of frontline AI users received no guidance about what to do with saved time, and more than half didn’t redirect it. Assign it to something specific, such as more client calls or a reduced backlog, or the savings won’t show up in results.
Step 6: Set rules for risk
Write these down before going live.
- Data. What can go into which tools? Are prompts used to train models? Where is the data stored?
- Review. Which outputs need a person’s approval?
- Accountability. Who owns the system and who handles complaints? A Canadian tribunal held Air Canada liable for wrong information from its website chatbot in 2024, rejecting the idea that the bot was separate from the company.
- Fairness and law. Hiring, credit, health and similar uses bring heavier obligations. In the EU, recruitment systems are high-risk under the AI Act, with obligations expected from December 2, 2027 according to Gibson Dunn’s summary of the provisional agreement. Check your own jurisdiction with counsel.
- Monitoring. Models and vendors change. Sample outputs regularly and keep track of quality.
Step 7: Decide, then scale carefully
At the end of the pilot, compare against your baseline, including licenses, integration, training and checking time. Then choose one of three options: scale it, change it or stop it. Stopping a pilot that didn’t work is a good use of the process, not an embarrassment.
Here’s a simple way to calculate the return, using invented numbers:
| Item | Annual figure |
|---|---|
| Hours saved (400 per month after review time) | 4,800 |
| Value at $45 per hour, if redeployed | $216,000 |
| Licenses | -$60,000 |
| Integration and maintenance | -$50,000 |
| Training and support | -$20,000 |
| Net benefit | $86,000 |
The biggest assumption in that table is that the freed hours become something valuable. If they don’t, the benefit could be close to zero, or negative.
When you scale, expect the human side to take the most effort: training, trust and changes in incentives. Our article on AI adoption focuses on that, and AI in business examples shows what results look like in practice. Small companies can use the simpler approach in AI tools for small business. For the wider picture, see AI and the future of work.
Frequently asked questions
How much does it cost to implement AI?
It runs from about $20 to $30 per user per month for off-the-shelf assistants to hundreds of thousands of dollars or more for custom systems. The price of the tool is often not the largest cost. Data preparation, integration, training and review staff can exceed it. Budget for those, and pilot before committing to multi-year deals.
Do we need to hire data scientists first?
Not for most first projects. Using commercial tools requires a clear business owner, someone technically capable and a policy on data. Specialists become more important when you build custom models, handle sensitive data at scale or need rigorous evaluation. Start by working out what problem you’re solving, then decide which skills you lack.
How long before we see results?
A focused pilot can show measurable results in two to three months. Wider returns usually take longer, because processes, training and systems need to change. Be suspicious of a promise of transformation within weeks, and equally wary of projects that run for a year without showing any measured improvement.
