Most SaaS teams have tried ChatGPT, a meeting note taker and an AI writing tool. Far fewer have an AI agent doing real, repeated work inside their sales, support or operations. The gap is rarely the technology. It is picking the right first task.
This guide gives you a simple way to score your tasks, the six that usually win, the ones to leave alone, and what it costs to run.
Score your tasks before you build anything
List the repetitive work your team does each week, then score each task from 1 to 5 on four questions:
| Question | Score high when | Why it matters |
|---|---|---|
| How often does it happen? | Daily, or dozens of times a week | Frequent tasks pay back the build cost fastest |
| Does it follow a pattern? | Similar inputs, similar steps, similar outputs | AI is good at patterns, weak at one-off judgment calls |
| Is the data available? | It lives in your CRM, inbox, docs or website | An agent can only work with data it can reach |
| How cheap is a mistake? | A person reviews it, or it is easy to undo | Low-risk tasks let you learn safely |
Add up the first three scores and subtract the fourth if mistakes would be costly. Your first project is the task at the top of the list, not the most impressive one.
Six tasks that usually win
1. Researching and scoring inbound leads
A form comes in. The agent visits the company website, checks size and industry, compares it with your ideal customer profile and writes a two-line summary with a fit score. Your sales team starts every call already knowing who they are talking to.
This is the most reliable first project we see, because the output is easy to check and a wrong score costs almost nothing.
2. Drafting first replies
Once a lead is scored, the agent drafts a reply in your voice, using your booking link and your answers to common questions. A person reads it and clicks send. Reply time drops from hours to minutes without anyone giving up control.
3. Answering questions from your own docs
Support and success teams answer the same questions every week. An assistant that only answers from your help center and policies, and shows which article each answer came from, handles the repeat questions and hands the rest to a person.
4. Researching prospects for outbound
Outbound works when the message is specific. An agent can watch funding news and hiring signals, read each prospect’s website and draft a short email around one real observation. Our own free website audit is a good example of the kind of research step an agent can run for every prospect: it checks speed, SEO and accessibility and returns findings you can mention in a first email.
5. Turning data into weekly reports
If someone spends Monday morning pulling numbers from three tools into a summary, an agent can do the pulling and the first draft of the summary. The person still decides what matters.
6. Extracting data from documents
Invoices, contracts, onboarding forms and applications often arrive as PDFs. Extracting the fields into your system is tedious for people and a good fit for AI, as long as someone reviews the unusual cases.
What to leave out of your first project
Some tasks look tempting but make poor first projects:
- Pricing and discount decisions. Mistakes cost real money and are hard to spot after the fact.
- Legal, medical or financial answers to customers. The cost of a confident wrong answer is too high.
- Anything you cannot easily check. If nobody can tell whether the output is right, you cannot improve it.
- Rare, high-stakes tasks. If it happens twice a quarter, the build will not pay back.
These can come later, once you have logs, review steps and a track record with simpler agents.
What it costs to run
People often expect AI model fees to be the big line item. For most internal tasks they are small. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and Claude Sonnet 5.5 at $2 and $10.
A lead summary might use around 3,000 input tokens and 500 output tokens. On Haiku 5.5 that is well under a tenth of a cent per lead. Even on Sonnet 5.5 it is about one cent. The real cost is designing the agent, connecting it to your tools and testing it on real cases, which is usually a fixed project fee.
“Pick the boring task your team complains about every week. That is almost always a better first agent than the exciting one.”
Keep a person where it matters
Decide up front which steps a person reviews: messages to customers, changes to important records, anything that spends money. Build those review steps in from the start, log what the agent did and why, and loosen the review only once you trust the agent on that task.
On data privacy, use business API plans rather than consumer chat apps. Anthropic states that by default it will not use inputs or outputs from its commercial products, including the API, to train its models. Send each task only the data it needs.
A simple first-month plan
- Week 1: list your repetitive tasks and score them with the table above.
- Week 2: prototype the top task on 20 to 50 real examples and check the output by hand.
- Week 3: connect it to your tools with a review step and logging.
- Week 4: run it for real, measure time saved and errors, and decide what to automate next.
If you also need tools connected without AI, such as forms to CRM or billing to Slack, that is classic workflow automation, and we compare the main platforms in Zapier vs Make vs n8n.
Sources
- Claude Haiku 5.5 pricing, Anthropic, October 2026
- Claude Sonnet 5.5 pricing, Anthropic, October 2026
- Is my data used for model training?, Anthropic Privacy Center