Short answer: AI automation works today for reading and sorting incoming documents and email, extracting data from them, drafting replies that a person approves, searching internal knowledge, summarising meetings and notes, and moving data between systems that do not talk to each other. It disappoints when it is asked to run a critical process with nobody checking, or when the task needs to be right every single time. Pick a first project by multiplying three things: how often the task happens, how well it tolerates an error, and how easily you can get at the data. Try an off-the-shelf tool first.
What “AI automation” means here
This guide is about language models applied to everyday office work: text, documents, email, notes and the records behind them. The model is one step in a workflow, usually a short one: something arrives, the model reads it and produces a structured result or a draft, code checks that result, and a person approves it or the system files it.
Much useful automation needs no AI at all. If a rule can express the task (“when an order is paid, create the invoice”), a conventional integration is cheaper, faster and exact. AI earns its place where the input is messy human language or documents that no rule can parse. The two combine well, which is why BBR’s software work for small businesses treats them as one toolbox.
Use cases that work today
| Use case | What it does | Why it works | Keep a person for |
|---|---|---|---|
| Document and email triage | Reads incoming mail and attachments; labels each by type, urgency and customer; routes it to the right person or queue | Classification is a strength of language models, and a wrong label is cheap to correct | Anything labelled uncertain; periodic spot checks |
| Data extraction | Pulls supplier, dates, amounts, line items or form fields from invoices, orders, applications and contracts into structured fields | Output can be validated by code: totals add up, dates are valid, the supplier exists | Approving each record before it is posted, at least at first |
| Drafting replies with human approval | Prepares a response to a routine enquiry using your policies, past answers and the customer’s record; staff edit and send | The slow part of replying is starting. Editing a good draft takes a fraction of the time, and nothing leaves without review | Every send |
| Internal knowledge search | Staff ask a question and get an answer with links to the procedure, policy or past project it came from | Answers show their sources, so staff can verify; saves interruptions to the two people who know everything | Keeping the underlying documents correct |
| Meeting and notes summarisation | Turns a transcript or rough notes into decisions, actions and owners; files them to the CRM or project tool | Low stakes, and participants can spot mistakes at once | A quick read before it is shared outside the company |
| Data entry between systems | Takes information from where it arrives (email, PDF, web form, spreadsheet) and creates or updates records where it belongs (accounting, CRM, booking, inventory) | Combines extraction with ordinary API integration; removes retyping, which is where human errors come from too | Exceptions, new suppliers or customers, amounts above a threshold |
The common thread: the model handles volume and the person keeps responsibility. Every item in the table has a natural review point, a way for code to check the output, or a low cost of being wrong.
What a good workflow looks like
- Trigger: an email, upload or form submission arrives.
- Read: the model classifies it and extracts fields into a fixed structure.
- Check: code validates the structure and the business rules. Does the purchase order number exist? Do the line items sum to the total?
- Review: a person sees the source document beside the extracted values, corrects anything wrong and approves. Items that failed a check are flagged.
- Act: the record is created in the target system through its API.
- Log: what came in, what the model produced, what the person changed. The corrections show you the accuracy and where to improve.
When the measured correction rate for a category stays low for some weeks, review for that category can drop to sampling. That decision should follow the numbers, not precede them.
Use cases that disappoint
| Idea | What goes wrong | What to do instead |
|---|---|---|
| A fully autonomous agent running a critical process, such as paying suppliers, issuing refunds or managing orders end to end | Errors compound across steps, the agent can be misled by text inside the documents it reads, and nobody notices until money or customers are affected | Let the agent prepare the action and a person approve it. Keep payment and deletion behind a human click |
| Anything needing perfect accuracy with no review: tax filings, payroll figures, legal deadlines, medical or safety information | Language models are very good and never perfect, and their mistakes look as confident as their correct answers | Use AI to prepare and to cross-check, with a qualified person signing off. Use deterministic software for the calculation itself |
| A public chatbot launched with no tested content and no route to a person | It answers confidently from outdated pages, and frustrated customers have nowhere to go | Clean the content first, test on real questions, and build the handoff. See our guide to what a support chatbot costs to build and run |
| Arithmetic, reconciliation and lookups done by the model itself | Models approximate. Sums across long tables and exact matches across lists are unreliable | Have the model extract the numbers; let code or a spreadsheet do the mathematics |
| Bulk AI-written marketing content published unedited | Generic text that readers and search engines both discount, with occasional invented facts | Use AI for outlines, first drafts and repurposing; a person with knowledge of the subject finishes it |
| “AI will sort out our messy data” | If records are incomplete, duplicated or trapped in a system with no export, the model has nothing reliable to work from | Fix access and data quality first. That work pays back with or without AI |
| Automating a task that happens ten times a month | Set-up and upkeep cost more than the time saved | Use a general AI assistant by hand for that task, and automate something frequent |
None of these is a permanent verdict on the technology. They are the places where, today, the gap between a convincing demonstration and a dependable process is widest. The reasoning behind agent limits is covered in more depth in our guide to building AI products and where agents fit.
How to pick a first project
List the repetitive, text-heavy tasks in the business, then score each on three factors from 1 to 3 and multiply.
| Factor | 1 | 2 | 3 |
|---|---|---|---|
| Volume | A few times a month | Several times a week | Many times a day, or hundreds a month |
| Tolerance for error | A mistake is costly and hard to notice | Mistakes are caught later at modest cost | A person reviews every output, or a mistake is harmless |
| Data access | Inputs are on paper, scattered or outdated; the target system has no API or import | Inputs are digital; the target system has an API or import that needs some work | Inputs arrive in one place digitally; the target system has a documented API |
Because the factors are multiplied, a score of 1 on any of them sinks the project, which is the intended effect. High volume does not rescue a task where errors are dangerous, and a perfect use case is worthless if the data cannot be reached.
A worked example
A 25-person wholesale distributor lists four candidates:
| Candidate | Volume | Error tolerance | Data access | Score |
|---|---|---|---|---|
| Entering supplier invoices from emailed PDFs into the accounting system, with approval before posting | 3 | 3 | 2 | 18 |
| Meeting summaries filed to the CRM | 2 | 3 | 3 | 18 |
| Internal knowledge search over procedures that live in old shared folders, many out of date | 2 | 3 | 1 | 6 |
| Automatic replies to customer complaints, sent without review | 2 | 1 | 3 | 6 |
Two candidates tie at 18. Meeting summaries are available in existing tools for a monthly fee, so that is switched on this week. Invoice entry needs a connection to the accounting system and a review screen, so it becomes the first custom project. Knowledge search waits until the procedures have been cleaned up and moved to one place. Complaint replies are reframed as drafts that a person approves, which lifts the error-tolerance score to 3 and the total to 18.
Before committing
- Measure the task today: items per month and minutes per item. Without a baseline you cannot show a saving.
- Collect 50 to 100 real examples, including the awkward ones.
- Name the person who will review outputs and who owns the process.
- Confirm access: can the inputs be read automatically, and does the target system accept records through an API or import?
- Write down what “good enough” means as a number, such as “nine in ten invoices need no correction”.
- Decide what happens to items the automation cannot handle. There will always be some.
Off-the-shelf or custom
Work through the options in this order, and stop at the first one that does the job.
| Option | Good for | Limits |
|---|---|---|
| 1. AI features in software you already use (email, office suite, helpdesk, CRM, accounting, meeting tools) | Summaries, drafting, meeting notes, receipt and invoice capture, search within that product | Works only inside that product; little control over behaviour |
| 2. A general AI assistant on a business plan, used by staff | Ad-hoc drafting, analysis, rewriting, low-volume tasks of every kind | Manual: someone copies in and out. Depends on each person’s habits |
| 3. A no-code automation platform with AI steps | Simple flows between popular cloud apps: “when an email arrives, extract these fields, add a row, notify the channel” | Awkward for branching logic, review screens and error handling; per-task fees grow with volume; hard to test. See no-code vs custom development for where the ceiling is |
| 4. A specialised product for one job (invoice capture, contract review, support chatbots) | A common task done well, with its own interface | Fits the vendor’s idea of the process, not yours; another subscription and another place where data lives |
| 5. Custom workflow | Tasks that span your particular systems, need your business rules, a review screen for your team, an audit trail, or run at a volume where per-task fees exceed the cost of owning the software | Up-front cost; needs a maintainer |
Signs that you have outgrown options 1 to 4: staff are copying data between an AI tool and another system all day; a no-code flow has grown to dozens of steps and fails silently; the process depends on a system with no ready-made connector, such as an industry-specific or older on-premise application; or you need to show who approved what. At that point a small custom tool, often an internal dashboard with a review queue and a few integrations, is simpler than the workaround it replaces. Connecting to the systems involved is standard API integration.
Privacy basics
A small business does not need a compliance department to use AI responsibly. It needs a short list of habits.
- Use business plans or API access, not personal accounts. Terms on training use, retention and confidentiality typically differ between consumer and business offerings. Read them, and note the date, because they change.
- Sign the data processing agreement the provider offers, and add the provider to your list of suppliers who handle personal data if your customers or regulators expect one.
- Send the minimum. An invoice-extraction step does not need the customer’s full history. Remove fields the task does not use.
- Keep a list of approved tools and a one-page staff policy: what may be pasted into which tool, and what may not (payment card numbers, health details, passwords, anything under a confidentiality agreement).
- Check sharing permissions before turning on AI search across company files. These tools surface whatever a user technically has access to, including the salary spreadsheet that was shared too widely three years ago.
- Keep a person in front of anything sent to customers, and do not let customers believe they are talking to a person when they are not.
- Decide how long prompts, outputs and logs are kept, and who can read them.
- Check sector rules. Health, legal, financial and children’s data carry extra duties in most jurisdictions.
- With custom software, keep provider API keys on the server, and log what was sent where.
Not legal advice. Data-protection duties depend on where you and your customers are and on the kind of data involved. If you handle sensitive records, speak to an adviser before sending them to any new supplier, AI or otherwise.
Rough cost bands
About these numbers. These are planning estimates to help you budget, not quotes, not survey data and not a BBR price list. Subscription prices change often, so check current vendor pricing. Custom figures follow the usual method of effort multiplied by rate, shown at two reference rates: $50 an hour (about $2,000 per person-week) and $150 an hour (about $6,000 per person-week). Per-token prices are placeholders.
| Route | Typical cost shape | Planning band |
|---|---|---|
| AI features in existing software | Included in the plan, or an add-on per user per month | From nothing extra to a few tens of dollars per user per month |
| General AI assistant, business plan | Per user per month | Commonly a few tens of dollars per user per month |
| No-code automation platform with AI steps | Monthly plan based on the number of tasks run, plus model usage | Tens to a few hundred dollars per month at small-business volumes |
| Specialised product | Per seat, per document or per resolution | Varies widely; multiply the unit fee by your real monthly volume before signing |
| Custom single workflow (for example invoice intake with a review screen and one integration) | 4–8 person-weeks including testing | $8,000–$16,000 at $50/hour; $24,000–$48,000 at $150/hour |
| Custom internal knowledge assistant with access rules | 12–18 person-weeks including testing | $24,000–$36,000 at $50/hour; $72,000–$108,000 at $150/hour |
| Upkeep of custom software | Monitoring, fixes, model version changes, small improvements | A common planning figure is 15–25% of the build cost per year |
Model usage is usually the small number
An illustration with placeholder prices of $1.00 per million input tokens and $4.00 per million output tokens. Extracting one invoice sends about 2,000 tokens and returns about 300:
- Input: 2,000 × $1.00 ÷ 1,000,000 = $0.0020
- Output: 300 × $4.00 ÷ 1,000,000 = $0.0012
- Per invoice: $0.0032. For 2,000 invoices a month: $6.40
Real prices differ by vendor and model, and scanned pages processed as images use more tokens. The order of magnitude is the point: for document workflows at small-business volume, the model bill is rarely what decides the project. The full method is in our guide to AI build and running costs.
A payback check
Suppose a task takes 4 minutes by hand and happens 1,500 times a month: 6,000 minutes, or 100 hours. With extraction and a review screen, each item takes 1 minute to check: 25 hours. The saving is 75 hours a month. At a placeholder staff cost of $30 an hour including overheads, that is $2,250 a month, so a $12,000 build pays for itself in a little over five months, before upkeep. Run the same sum with your own volume, minutes and wage. If the payback is longer than about eighteen months, choose a cheaper route or a different task.
A four-week plan for a first project
- Week 1: list candidate tasks, score them, pick one. Measure the baseline. Collect examples. Check what your existing tools already offer.
- Week 2: try the task by hand with a general AI assistant on twenty real examples. If quality is poor here, automation will not improve it. If it is good, you have your prompt and your first test set.
- Week 3: build the simplest workable version: a built-in feature, a no-code flow, or a custom prototype if the systems involved require it. Keep a person reviewing every item.
- Week 4: run it on live work alongside the old process. Record the correction rate and time per item. Decide whether to keep it, improve it or stop.
Then repeat with the next task on the list. Several small automations, each measured, do more for a small company than one ambitious programme. If the task you have chosen needs custom software, BBR’s AI development service begins with the same steps: the task, real examples and a prototype that shows whether the quality is there, built on hosted models from established vendors and not on models we train ourselves.