Practical notes on AI, product, design and shipping software — written by the people who build it.
An approval button is only the start. Define what the person sees, which proposal they approve and what happens when the facts change or execution fails.
Measure the cost of a completed task before changing your model. Compare caching, routing and retries without losing the quality controls that matter.
Watch people try the risky journey before paying to build it. A fictional booking example shows how to separate observations, interpretations and design changes.
A model leaderboard cannot choose your production system. Define acceptance rules, compare the same cases and price the work that actually passes.
A no-training statement does not answer where prompts, files or logs are stored. Map the copies and test deletion before connecting business data.
Governance becomes useful when someone can explain who approved a workflow, what it may do and how to stop it. Start with one filled system record.
A useful audit connects a concrete observation to a consequence and a verifiable fix. This fictional booking example shows the difference.
A document can be read correctly and still produce an unusable record. Follow three delivery notes through mapping, validation and exception handling.
Replace the fictional inputs with your scoped estimates. The calculator separates the build, contingency and the first year of operation.
A good reply is not proof that an action happened. Use structured task evidence to identify false success, unauthorised actions and duplicate execution.
Extraction can be correct while the invoice still fails approval. Follow a partial delivery through matching, review and payment controls.
A delivery lookup and a disputed delivery need different support. Map the decisions, design the handoff and measure outcomes beyond conversations that simply end.
An extra retrieval step can help without needing an agent. A runnable fictional fixture shows how to choose a fair baseline before testing a planner.
A wrong answer can be a retrieval, behaviour or permission failure. Diagnose it before buying a training pipeline, using a fictional policy example and a reusable failure log.
Build a chatbot budget around the whole support task: setup, software usage, knowledge upkeep and the work left for people. Published prices are kept separate from fictional workload assumptions.
Compare what support-chatbot projects actually demonstrate: useful answers, connected systems and the route to a person. Five suppliers, clear evidence limits and a practical demonstration sheet.
A prototype, live pilot and platform rebuild answer different questions. Compare six teams through published projects, then use a shared first-release brief to assess the proposals.
Published consulting offers buy different things: advisory time, a leadership review or technical discovery. Compare the included investigation and deliverables before comparing totals.
Choose an AI consultant around the decision you need to make. This disclosed shortlist distinguishes delivered work, research and advisory offers, with a brief for comparing suppliers.
Web-app estimates depend on permissions, data sources and transactions, not just screens. Compare three fictional delivery budgets with Pixelfield’s published price guidance and a first-year operating-cost example.
Test field evidence, routing and CRM writes separately. This synthetic lead pack includes missing budgets, support requests, contradictory inputs and duplicate events, with expected decisions rather than claimed model performance.
An MVP estimate should price a testable first release, including the work your team will perform manually. The fictional £36,800 model makes scope, assumptions and exclusions visible; it is not a Pixelfield tariff.
Pixelfield production agent builds typically start at £30,000. A useful estimate also prices permissions, integrations, evaluation and operating effort. The worked budgets here are explicitly fictional assumptions, not a quote.
Start with one process and its exceptions. This four-provider comparison distinguishes operational automation, AI task evaluation and governed data access, then shows what to request in a workflow proposal.
Choose an AI supplier by relevant project evidence and the system you need delivered. This disclosed six-company shortlist separates custom development, enterprise integration and platform adoption, with a reusable buying brief.
Subtract routine review, exception handling and running costs before valuing time saved. This editable calculator estimates capacity-based ROI and break-even; it also shows when an automation loses value.
Compare PWA and native app options around offline work, push notifications, hardware and distribution. Includes a practical acceptance-test checklist.
Plan an AI development budget with published UK prices, a worked build and operating cost example, and a worksheet for comparing supplier proposals.
An AI proof of concept needs an agreed investment decision. Follow a fictional product-question assistant through evidence, acceptance gates and a revise-and-retest result.
A chat interface can hide a fixed workflow or a model-led agent. Compare three designs for one appointment-change task, with permission boundaries and eight acceptance checks.
