Service · Automation
Somewhere in your business, a capable person is spending their Tuesday copying numbers from one screen into another. We find that work and give it to software — reliably, with an audit trail, and with a human still in charge of anything that matters.
What is AI workflow automation?
AI workflow automation is software that carries a repetitive business process from start to finish without a person doing each step — receiving a trigger, reading and understanding messy inputs such as scanned invoices or WhatsApp messages, updating the right systems, and escalating only genuine exceptions to a human. It differs from older automation in that it copes with unstructured, inconsistent input rather than breaking on it.
The honest test of whether a process is worth automating is not how advanced it sounds. It is this: does someone repeat it at least weekly, does it follow rules more than judgement, and does getting it wrong cost real money? If the answer to all three is yes, automation will almost certainly pay for itself inside a year — often inside a quarter.
We start by watching the work as it is actually done, which is rarely how the procedure manual describes it. The undocumented WhatsApp group where approvals really happen, the spreadsheet one person maintains privately, the workaround everyone uses because the system refuses a valid entry — those are where the time goes, and they never appear on a process map.
Signs you need this
In practice
These are the workflows we are asked for most often in Kenya, roughly in order of how quickly they pay for themselves.
Pull statements automatically through the Daraja API and bank feeds, match payments to invoices with fuzzy name and reference matching, post the confident ones, and queue only real exceptions for a human. This is the single highest-return automation for most Kenyan businesses.
Supplier invoices, delivery notes, LPOs and receipts — scanned, photographed or emailed — read into structured data, validated against the purchase order, and pushed into your accounting system. Including the ones photographed badly on a phone at a loading bay.
Quotes, purchase requests, leave, credit notes and discounts routed to the right approver by value and category, chased automatically, escalated when they stall, and logged permanently. No more “I never saw that message”.
The daily sales summary, the weekly stock position, the monthly management pack — assembled from source systems, checked for anomalies, and delivered to WhatsApp, email or a dashboard before anyone asks for them.
Keeping your e-commerce store, accounting package, inventory system and CRM telling the same story, continuously, instead of someone reconciling them on a Friday afternoon.
Collecting customer or member documents, extracting the fields, running validation and sanction checks, flagging what is missing, and assembling a clean file for review. Common for SACCOs, insurers and lenders.
The same answer, two ways
How a workflow automation is actually built
Think of it as hiring a very fast, very literal assistant who never sleeps.
First we sit with the people who do the work and watch them do it. We write down every step, including the annoying ones they have stopped noticing — the login that always fails first time, the file that has to be renamed, the colleague who has to be reminded on WhatsApp every Thursday.
Then we pick the steps a machine can do safely and build software that does them. The software watches for the trigger (a payment lands, an email arrives), does the work, and puts anything it is unsure about into a short list for a person to look at. That list is usually a tiny fraction of what the person handled before.
We test it beside the existing process for a week or two — both running at once — so you can see that the machine agrees with your team before you trust it alone. Then we switch the manual version off, train everyone on what to do when something odd happens, and write it all down in plain English.
Nothing disappears into a black box. You can always see what was done, when, and why. And if the automation ever needs to be switched off, it switches off, and your team carries on as before.
Event-driven orchestration over your systems of record, with typed extraction, confidence gating and idempotent writes.
Discovery. Process mapping with cycle-time, touch-count and rework-rate baselines. Candidate scoring against value, effort and risk. Data-flow mapping including where personal data enters and leaves, which drives the Data Protection Act analysis.
Ingestion. Webhooks where the source supports them (Daraja C2B and B2C callbacks, WhatsApp Cloud API, Stripe-style notifications), polling with cursors and watermarks where it does not, IMAP for email-borne documents, and watched folders for scanner output. Everything lands in a durable queue first so nothing is lost when a downstream system is unavailable.
Extraction. For structured input, plain parsers. For documents, layout-aware OCR followed by an LLM extraction pass constrained to a JSON schema, with per-field confidence, plus deterministic validators — checksum on KRA PINs, arithmetic checks on invoice totals, date-range sanity, supplier existence against the master list.
Decisioning. Confidence thresholds tuned against a labelled historical set rather than guessed. Above threshold and validators pass: auto-post. Below: exception queue with the specific reason and the source document rendered alongside. Money-moving actions always require human approval regardless of confidence.
Writes. Idempotency keys on every mutating call so retries cannot double-post, compensating actions for rollback, and reconciliation jobs that verify the target system reflects what we believe we wrote.
Operations. Structured logs, per-step metrics, dead-letter queues with alerting, and a shadow-run period where the automation processes live traffic without writing, so its output can be diffed against the human process before cutover.
How cleanly common tasks automate. Higher means less human judgement is required — and therefore a faster, cheaper and lower-risk first project.
What you get
Not a report about automation. The automation, running, with the documentation your team needs to own it.
Typical engagement
Every quote is fixed before work starts. If we scoped it wrong, that is our risk, not yours. See how pricing works.
Typical stack
We choose tools your team can maintain, not the ones that make us look clever.
Questions
AI workflow automation is the use of software — a combination of fixed rules and artificial intelligence models — to carry out a repetitive business process from end to end without a person doing each step by hand. A typical automated workflow receives a trigger (an email arrives, a payment lands, a form is submitted), reads and understands the content, checks or updates the relevant systems, and either completes the task or routes an exception to a human. The AI part matters when the input is messy — a scanned invoice, a WhatsApp message, a handwritten delivery note — because traditional rule-based automation breaks on anything it has not seen before.
Classic robotic process automation (RPA) drives the user interface: it clicks buttons and types into screens exactly as a person would, and it snaps the moment a layout changes or an unexpected value appears. We prefer to work at the API and data layer where possible, which is far more robust, and to use language models only where genuine interpretation is required. Where a system has no API and screen automation is the only option, we will use it — but we will tell you honestly that it is the fragile part of the design and price the maintenance accordingly.
The one that is high volume, highly repetitive, low in judgement, and already causing pain. In Kenyan businesses that is very often payment reconciliation, invoice or delivery-note capture, or the daily and weekly report that someone assembles by hand. We score candidates on value, effort and risk during the audit, and deliberately start with something small enough to ship inside six weeks — early visible wins are what unlock budget for the bigger ones.
It is designed to. Every workflow we build has a confidence threshold: above it the system acts, below it the item goes into an exception queue for a human, with the reason attached. Anything that moves money or creates a legal commitment always requires human approval regardless of confidence. Every action is logged with its inputs and outputs, so a wrong decision can be traced and reversed rather than quietly compounding.
Almost never. We build around what you already run — QuickBooks, Sage, Odoo, Zoho, Excel, Google Workspace, a custom system a developer built in 2014, or all of the above at once. Ripping out working systems is expensive, slow and politically painful; connecting them is usually cheap and fast. If a system genuinely cannot be integrated, we will say so and price the workaround honestly.
A focused first automation typically takes two to six weeks from kickoff to production, assuming we can get access to the systems involved in the first week. Access and permissions are almost always the bottleneck, not the engineering. Larger programmes covering several processes run three to six months, but we deliver them as a sequence of small live releases rather than one big launch.
Keep exploring
Most clients combine two or three of these.
A WhatsApp and website assistant that answers customers in English and Kiswahili, qualifies leads, checks order status and escalates to a human when it should.
An honest map of where AI will make you money, where it will waste your money, and what to do first — costed, sequenced and tied to your numbers.
When off-the-shelf will not fit: document intelligence over your own files, forecasting on your own history, and integrations into the systems you already run.
Start here
Book a free 30-minute call. We will map one process end to end, tell you honestly whether AI is the right answer, and put a number on what fixing it is worth. No slide deck, no jargon.
Nairobi-based · We reply the same working day · English & Kiswahili
Talk to a human