Which business task should you automate first?
Choose a task that happens often, follows a recognisable process and produces an output someone can assess. Sorting incoming enquiries, preparing a proposal draft and matching expense documents are reasonable candidates. A broad instruction such as “make our company use AI” gives neither the team nor the supplier a reliable finish line.
Talk to the people doing the work before selecting a tool. Ask them to walk through a normal item and a difficult one. Note where they copy information, wait for approval, correct mistakes or search for a document. Those steps often reveal whether the problem needs AI, a conventional integration or a better form.
For a B2B company in Jordan, Saudi Arabia or the UAE, language is part of this assessment. An English demonstration does not establish that the same workflow can read an Arabic quotation, preserve a company name or distinguish a tender from a supplier introduction. Include the material your team actually receives.
Write a one-page pilot brief
The brief should name the process, its owner, the inputs, the proposed output and the person who approves the result. It should also say what the pilot will leave outside its scope. A bounded project is easier to assess and price than a promise to automate an entire department.
For example: “Receive service enquiries from two shared inboxes, identify the service requested, suggest an owner and prepare a daily review list.” That is specific enough to discuss data access and acceptance tests. Sending replies, changing CRM ownership and submitting commercial offers can remain separate decisions.
List the systems involved without assuming every connection already exists. An Outlook mailbox, a spreadsheet export and a CRM may each require different permissions. Include access arrangements and the agreed retention period in the implementation plan. Keep the business owner involved when a technical shortcut would change the approved scope.
Measure the complete current process
Count time spent collecting information, checking it, handling exceptions and producing the final output. Measuring only the fastest part of the process creates an unrealistic comparison. It is also important to record error types, because a quick workflow that requires substantial correction may not save useful time.
Use a representative sample rather than the easiest five examples. Include missing attachments, duplicate submissions, unclear wording and different document formats. Record which cases required an experienced employee. This gives the pilot a more honest baseline than a general estimate of how busy the team feels.
A simple spreadsheet can contain item ID, category, handling minutes, correction minutes and outcome. It does not need private customer information. Agree what can be anonymised before copying examples into a test environment. A supplier should be able to explain which data is necessary and which can be omitted.
Decide whether AI is needed at each step
Use ordinary rules for exact calculations, mandatory fields, duplicate identifiers and permission checks. AI is useful where language or unstructured information makes fixed rules difficult. A workflow might use a model to classify an enquiry, then conventional code to assign a queue and calculate a deadline.
Do not ask a language model to invent a price or approve an expense because it can write a convincing explanation. Keep those decisions attached to a controlled source and an authorised person. If the source is absent, the system should raise an exception rather than fill the gap with a plausible value.
This distinction also improves maintenance. When a department changes an approval threshold, a visible rule is usually easier to update and test than a large prompt containing many hidden exceptions. Our custom software development service covers these integrations and business rules alongside the AI work.
Build a test set before the demonstration
Agree the examples and expected outcomes before reviewing the supplier’s best demonstration. Use some records for development and hold others back for evaluation. The held-back set should resemble real work, including Arabic and English cases, rather than a collection of questions written to match the prompt.
For classification, record correct category, uncertain category and consequential error separately. Mistaking an ordinary newsletter for an enquiry is inconvenient; overlooking a tender deadline may have a different business impact. Decide which errors should stop deployment and which should enter a review queue.
For document generation, check required sections, factual claims, numbers, dates and source references. For a knowledge assistant, check whether the answer is supported and whether the user was entitled to see the source. Testing the output alone can miss a permissions failure elsewhere in the workflow.
Put human review where it changes the outcome
Review should have a named owner and a visible place in the process. “A human stays involved” is too vague if no one knows who receives uncertain cases or how quickly they should act. Define what the reviewer sees, what they can change and how the decision is recorded.
Start with a read-only or draft-producing workflow where that suits the task. Once it performs consistently, discuss whether selected actions can be automated under explicit rules. Outbound replies, payments and commitments deserve separate approval and testing; a successful summary does not prove those actions are safe.
Our AI solutions and automation work includes enquiry, document and finance workflows. The relevant question is how a particular process should run, not whether every possible action should happen without an employee.
Compare useful time and full operating cost
Calculate time after automation as preparation plus processing supervision, review, corrections and exception handling. Include model usage, hosting, integrations and support in operating cost. Keep the one-off implementation fee separate so the buyer can compare recurring work value with the initial investment.
An illustrative example: 200 items taking six minutes each require 20 staff hours. If preparation and review together take eight hours after implementation, the model releases 12 hours. At an assumed labour-time value of SAR 80 per hour, that is SAR 960 of capacity before software and support costs. These are example assumptions, not a client result or cash saving.
Released time does not necessarily reduce payroll. It may let the same team respond earlier or clear a backlog. Agree which outcome matters, then measure it. The invoice reconciliation guide explains how to include review effort in a document-heavy workflow.
What should be handed over?
Ask for a process map, access inventory, configuration notes, test results and an operating guide. The guide should explain how to rerun a failed item, correct a record, stop the workflow and contact the responsible person. A video of the happy path is useful, but it does not replace these instructions.
Also agree ownership of accounts, data exports and generated documents. A business should know what happens if an API changes, a provider reaches its quota or the person who built the workflow is unavailable. Include a small recovery exercise before calling the pilot ready for everyday use.
The next step may be a larger deployment, a narrower pilot or no further work. All three can be sensible outcomes. A pilot earns its value by making the decision clearer, even when it shows that a simple non-AI integration would solve the problem more reliably.
Questions buyers often ask
Do we need to train our own model?
Usually not for a first workflow. Configuration, retrieval from approved documents and conventional integrations may be sufficient. Fine-tuning becomes a separate technical decision when testing identifies a specific need and suitable data exists. It does not replace accurate source records or access controls.
Can a small team run the pilot?
Yes, if someone owns the process and reviewers have time to assess results. A small team should begin with a smaller workload and fewer connected systems. Availability for corrections matters more than the number of people shown on a project organisation chart.
How do we choose the next project?
Use what the first pilot teaches about data quality, review load and maintenance. A bilingual proposal workflow or an Arabic knowledge assistant may be a useful next step if it shares tested foundations and has a clear business owner.
Sources and practical next steps
The NIST AI Risk Management Framework provides a structured way to govern, map, measure and manage AI risks. It is a useful reference for agreeing responsibilities and evaluation, rather than a certificate that a particular implementation is reliable.
Bring a sample process, an approximate monthly workload and a few cleared examples to the first discussion. Markenity can help define the pilot, build the required connections and agree how your team will judge the result.