AI Agents for Business: 12 Workflows Worth Automating, and the Ones That Aren't

Most owners have heard that AI agents can take work off their team. Their questions are plainer. Which jobs in my firm? What does it cost? What happens when it gets something wrong?

The first thing I ask about any task is how often it happens. A daily task can repay its build within months; one that comes up twice a month usually never does.

Below are twelve workflows for a small or mid-sized firm: what the agent does, what a person still approves, the time saved and the running cost. Then the ones I would leave alone. I build these in a fixed-price AI Agents & Automation Sprint: one to three workflows in 30 days, handed over into your own accounts.

The short version

• An agent gathers, drafts and checks. A person approves anything that leaves the office, spends money or makes a promise.
• Volume decides. Daily tasks repay their build fastest; monthly ones rarely do.
• Running costs are small next to the build: by my estimates, from almost nothing to €120$135 a month per workflow.
• Start at the front door: enquiries, onboarding, follow-ups after calls.
• Leave rare tasks, judgement calls and changing processes alone.

What is an AI agent, in plain terms?

An AI agent is software that reads, decides the next small step and drafts, using a language model: the engine behind ChatGPT or Claude. Think of a fast new assistant whose work is checked before it goes out.

Anthropic, the company behind Claude, separates fixed workflows from agents that choose their own steps, in its guide to building effective agents. In a small firm, what pays back is usually a fixed routine with an agent for the open-ended part.

Many agent projects will not last. In June 2025 Gartner predicted that over 40% of them will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. The answers: do the sums first, cap the spending, keep a person at the end.

Which jobs can an agent take over, and what do you still approve?

The gathering, drafting and checking. You keep the decision and the send button. Seven of the twelve below come from systems I have built, for a law firm and for my own products; five are common requests.

With clients and prospects

1. Replies to new enquiries. The agent reads the enquiry and the sender’s website, checks your CRM (the system holding your client records) and drafts a reply and a call brief. You edit and send; it never emails a prospect itself. Specific replies matter: Gartner found that 73% of B2B buyers actively avoid suppliers who send irrelevant outreach.

2. Client onboarding brief. The agent researches a new client’s website and pre-fills your intake questions, so the client corrects rather than writes. The client confirms each answer; you approve the brief. My own intake form, BriefDesk (not public yet), takes answers drafted by any AI and saves each as the client types. What it proves: projects start with the content in hand.

3. Proposal and tender drafts. The agent drafts each answer from past proposals, with its source, and flags what it could not answer. You set scope, price and commitments; it never invents a price or a reference.

4. Meeting notes to follow-up. From the call transcript, the agent writes a summary, proposed CRM updates, tasks with owners and a follow-up email. Whoever ran the call confirms and sends.

5. Social posts approved by phone. The agent drafts a week of posts per platform in the brand’s voice; the client approves them from a phone. This is how Cuckoo, my own product in pilot, works (publishing awaits platform approvals). What it proves: approvals move from email threads to a phone.

In the back office

6. Emailed request to researched draft. Staff email a one-line request, such as a company name, to a set address. The agent checks official registers and public warning lists, quotes each source word for word and drafts the text. A professional approves it before it is published.

7. Fact check against sources. The agent splits a draft into single claims and marks each as supported, partly supported or unsupported. Banned phrases fail outright; for a law firm, anything that sounds like a promised outcome. Unsupported claims are cut or held for an editor.

8. Staff questions from your own documents. The agent searches policies, manuals and past projects, answers only from what it finds and names the document, or says “not found”. It can read but not change anything, and each person sees only what their role allows.

9. Supplier invoices into the books. The agent reads each invoice from a mailbox, pulls out supplier, amounts, VAT and dates, and prepares a draft entry. Mismatches go to a queue. Your bookkeeper posts the entries.

Keeping watch

10. Daily approval email. One morning email lists every item that passed its checks, with approve and preview links. Each link works once, for one item, then expires. It makes workflows 5, 6 and 7 safe.

11. AI visitor monitor. The agent counts visits from AI crawlers, the programs AI assistants send to read websites, and alerts you when they stop coming. It changes nothing. If you suspect a block, read my piece on Cloudflare blocking AI crawlers.

12. Morning report with a spending stop. Each morning: what ran, what failed, what it cost. At a monthly cap it stops spending; only a person can raise the cap. Filyhost, my file-delivery service, pauses new sign-ups at its monthly budget while existing links keep working. What it proves: a budget limit need not break the service.

How do you know whether an agent pays for itself?

Compare the hours it saves, valued at what they cost you, with the cost to build and run it. If three months of savings cover the build, it is worth doing.

• An hour of staff time. I use €40$45, wages plus overheads. Eurostat put the euro-area average at €38.2 per hour in 2025; senior time costs more.
• Monthly savings: items per month, times minutes saved per item, times that hourly cost.
• Running cost: AI usage at the provider’s published prices (Anthropic’s, for Claude), plus hosting and email. My ranges are generous; one draft usually costs cents.
• The bar. The sprint costs from €7,500from $8,300 for up to three workflows, so each carries a third. A workflow passes when three months of savings, minus running costs, cover that third. Build only one and the volumes roughly triple.

These are my estimates; your volumes decide. I count hours only: won deals are real, but too easy to inflate.

A worked example: a hypothetical accounting firm

Picture a 12-person accounting firm that gets about 50 new enquiries a month. A partner spends around 40 minutes on each: reading it, looking the company up, checking past dealings, writing a reply.

With workflow 1, the agent does the reading and looking up. The partner spends about 10 minutes checking the draft. That saves about 25 hours a month, more than three working days.

Even at €40$45 an hour, likely well below what a partner’s hour costs the firm, that time is worth about €1,000$1,125 a month. Running costs would be around €30$35 a month. As one of three workflows in a sprint, it would typically repay its third of the price within 3 months. Clients would notice first: a same-day reply that speaks to their situation.

The same firm’s supplier invoices tell the opposite story. About 80 a month, each 4 minutes quicker, comes to roughly 5 hours, worth around €213$240 a month. It would take about 14 months to repay its share. In a case like this I would wait until volumes grow.

What each workflow saves and costs

The last column is the monthly volume at which a workflow repays its third of the sprint price within 90 days.

Workflow	Time saved, after checking	Running cost a month	Worth it from
1. Replies to new enquiries	30 min per enquiry	€10–€30$10–$35	44 enquiries a month
2. Client onboarding brief	3 h per new client	€5–€15$5–$15	8 new clients a month
3. Proposal and tender drafts	4 h per proposal	€10–€30$10–$35	6 proposals a month
4. Meeting notes to follow-up	15 min per call	€10–€30$10–$35	87 calls a month
5. Social posts approved by phone	2 h per brand per week	€15–€45$15–$50	3 brands
6. Emailed request to researched draft	1 h per request	€20–€60$20–$65	23 requests a month
7. Fact check against sources	45 min per document	€15–€40$15–$45	30 documents a month
8. Staff questions from your documents	15 min per person per day	€30–€120$35–$135	5 staff
9. Supplier invoices into the books	4 min per invoice	€5–€25$5–$30	322 invoices a month
10. Daily approval email	10 min per item	€0–€10$0–$10	127 items a month
11. AI visitor monitor	1 h per website per week	€0–€5$0–$5	6 websites
12. Morning report with a spending stop	10 min per system per day	€0–€5$0–$5	7 systems

The approval email and the morning report rarely pay back on time alone; they make the others safe. Invoice intake needs hundreds of invoices a month.

Which workflows are not worth automating?

Rare tasks, judgement calls, changing processes, and work where checking takes as long as doing.

• Rare work. A quarterly board pack happens four times a year; no build repays itself at that volume.
• Judgement calls. Whether to take a case, what to charge, what to tell a patient. An agent can prepare the file; a qualified person decides.
• Processes in flux. If the steps change every month, the agent is wrong every month. Fix the process first.
• Hard-to-check work. If your bookkeeper must redo every figure to trust it, nothing is saved.
• Two-step jobs. Copying a web form into a spreadsheet needs a simple automation tool, not an agent. Set it up yourself and keep the money.
• A chatbot for its own sake. One that answers buyers unchecked can promise what you do not offer.

What does this look like in a real business?

The clearest example is the content engine I built for a German law firm, Corvane Legal in my case study. It combines workflows 6, 7, 10, 11 and 12.

Observed on 5 October 2026: 1,100+ pages live on the firm’s own domain, a check for new work every 5 min and 649 automated tests. Limiting rewrites per day holds AI spend to ~€33/mo, under a hard €100 monthly stop. In Sept 2026 it moved into the firm’s own Cloudflare account. The keys, the logs and the bill are theirs.

Rejections matter as much as output. For the same firm, agents proposed 1,175 links between articles drawn from one Word file. Automatic checks threw out 72 made-up ones: link text that was not there, or pages that do not exist. A relevance check removed 179 more; 924 went live. That is what checks are for.

How do you stay in control of an agent?

Give it as little access as the job needs, cap its spending, log everything and keep a person where it matters. OWASP, a security project, calls the main risk excessive agency: an agent with too many functions, too many permissions or too much freedom.

• Spending caps in two places: at the AI provider, and a stop in the workflow that triggers first.
• Your accounts from day one. AI keys, hosting and domains are yours; I am a member you can remove.
• Least access: draft-only in your CRM and accounting system, read-only for document search.
• A log of every action and every cent, including who approved it.
• Stop loudly. If a source changes, the workflow stops and emails you instead of guessing.
• Pilot mode first. Every item is approved by hand until its record earns more freedom.

An agent I build never moves money, changes bank details or accepts terms. It never deletes data (it archives), writes reviews or testimonials, approves its own work or raises its own limits. It sends nothing in your name without approval, unless you have moved that workflow to automatic in writing.

Which workflow should you automate first?

The one that happens most often and is easiest to check, usually at the front door. Two or more “no” answers to these five questions mean wait.

• Does it happen at least weekly, near the volume in the table?
• Can a person check the result in under a minute?
• Is the information already in a system an agent can read, such as your inbox or CRM?
• Is a wrong draft cheap, because a person checks it first?
• Can you name the person who approves?

If your enquiries arrive through your website, check first that AI assistants can read it. My piece on what an AI-ready website needs covers that.

How much does it cost to have one built?

The AI Agents & Automation Sprint costs from €7,500from $8,300, typically €7.5k–15k$8.5k–16.5k, at a fixed price over 30 days. It starts with a workflow audit: your repeat work ranked by hours saved, risk and access. Then I build one to three workflows, test them on your real data and hand over the code, a runbook and every account. You pay running costs directly to the providers; I add no markup.

It is not for you if you want an agent that acts unchecked, if your volumes sit far below the table, or if the process still changes every month.

What you can do this week

• List and count the repeat jobs (45 minutes). With your team leads, note every task that came up more than ten times last month. Count from your inbox, CRM or accounting system, not from memory.
• Time the top three (about 10 minutes each, across the week), including the checking.
• Do the sums (15 minutes): volume, times the minutes a draft would save, times your hourly cost. Compare the result with the table.
• Name the approver (5 minutes): one person per workflow. No name, no agent.
• Check where the inputs live (15 minutes). If the information sits on paper or in one person’s head, that comes first.
• List what is off-limits (10 minutes): salaries, HR files, client secrets.

For your web developer

• Spend: a monthly limit per Claude API workspace (not possible on the Default one) plus an in-code breaker that trips first. Every provider call is logged with its cost.
• Access: each platform’s own OAuth sign-in, never stored passwords; draft-only scopes; role-based, read-only retrieval behind single sign-on. Approvals are single-use, expiring, per-item links.
• MCP (an open standard that connects agents to tools): Vowko, my marketplace for people who already trust each other, gives members’ agents 9 scoped tools with spending caps and an audit log. What it proves: an agent can use your data without faking trust.
• Tests on captured real cases, so a model update cannot quietly change behaviour.

Send me the workflow you would automate first and its monthly volume, and I’ll reply with the payback maths before we talk price.

Call it AI or Super Intelligence (SI). Your buyers already ask it before they ask you.