Question 1

GA4 is free — why pay for help with it?

Accepted Answer

Because the default configuration answers almost nothing useful. Value comes from correct event and conversion design, server-side reliability, and consent handling that stands up to scrutiny. Misconfigured analytics is worse than none: it produces confident, wrong decisions.

Question 2

How do you handle consent and privacy?

Accepted Answer

Consent mode configured properly, with clear treatment of what is measured before consent and what is modelled afterwards. We report the distinction rather than presenting modelled figures as observed ones.

Question 3

Can it show us actual revenue attribution?

Accepted Answer

Partly, and honestly. GA4 gives directional attribution; joined with CRM and platform data in Power BI it gets considerably closer to truth. Anyone promising exact multi-touch attribution is overselling the medium.

Question 4

Why Power BI rather than the reporting in our tools?

Accepted Answer

Because platform-native reporting only ever sees its own data. The decisions that matter need marketing, commerce and operational figures side by side, with agreed definitions. That consolidation is the whole point.

Question 5

How do we know the numbers are right?

Accepted Answer

Because the definitions are agreed with the people who use the report before anything is built — what counts as a lead, when revenue is recognised, how returns are treated — and anything modelled or estimated is labelled as such on the dashboard.

Question 6

Can different people see different data?

Accepted Answer

Yes. Row-level security means a regional manager sees their region and an executive sees everything, from one report rather than several maintained copies.

Question 7

Is Dynamics 365 worth it for a mid-sized business?

Accepted Answer

It is when the alternative is a spreadsheet layer between systems. The value is a single record of the customer across sales, service and finance. If your processes are simple and stable, a lighter CRM may serve you better and we will say so.

Question 8

How long does implementation take?

Accepted Answer

It depends far more on your process clarity than on the software. The work that determines the timeline is agreeing definitions and data ownership. We scope in stages so something useful ships before the whole programme completes.

Question 9

Can it talk to our website and marketing tools?

Accepted Answer

Yes. Connecting Dynamics to web forms, commerce and GA4 so a lead can be traced to revenue is one of the more common integration jobs we do — and one of the more valuable.

Question 10

Why Shopify Plus over a custom build?

Accepted Answer

Because payments, checkout, PCI scope and fraud handling are solved problems, and rebuilding them rarely repays the cost. We spend the budget on the parts that differentiate you — merchandising, integration and the post-purchase experience.

Question 11

Can it integrate with our ERP and inventory?

Accepted Answer

Yes, and that integration is usually where commerce projects succeed or fail. Inventory truth, pricing rules and order flow need to reconcile with the back office, and we build that with idempotent, monitored connectors rather than nightly guesswork.

Question 12

What about the platform limits?

Accepted Answer

Real, and worth knowing before you commit. Checkout customisation and certain B2B pricing models have boundaries. We map those against your requirements up front rather than discovering them mid-build.

Question 13

Why AWS specifically?

Accepted Answer

Depth of managed services, mature compliance tooling and the breadth of the local talent pool. For payments and regulated work in particular, the audit and encryption primitives are well-established and well-documented, which shortens every compliance conversation.

Question 14

Are we locked in?

Accepted Answer

Partly, and we are direct about it. Managed services trade portability for reliability and speed. We keep application logic portable and are explicit about which components are deliberately AWS-specific, so the trade is a decision rather than a surprise.

Question 15

How do you keep cost under control?

Accepted Answer

Right-sizing, autoscaling, lifecycle policies on storage and tagging so spend is attributable to a service. Reviewed on a cadence — cloud cost creeps upward by default unless somebody is looking at it.

Question 16

Why Node for the back end?

Accepted Answer

Shared language across the stack reduces context-switching and duplicated logic, and the ecosystem for API and integration work is deep. For I/O-heavy workloads — which most integration work is — it performs well and staffs easily.

Question 17

How do you handle long-running or heavy jobs?

Accepted Answer

Off the request path. Queues, workers and scheduled processes with retry and dead-letter handling, so a slow third-party system degrades one job rather than the whole application.

Question 18

Is it suitable for regulated workloads?

Accepted Answer

Yes, with the surrounding discipline: dependency scanning, pinned versions, secrets management, structured logging and least-privilege access. The language matters less than the controls around it.

Question 19

Does this help or hurt SEO?

Accepted Answer

It helps, when configured correctly — and hurts badly when it is not. Server rendering is the difference between crawlers seeing your content and seeing an empty shell. We have recovered a client from exactly that failure, so we treat it as a launch gate rather than an optimisation.

Question 20

How do you handle performance?

Accepted Answer

Rendering strategy per route rather than one setting for the whole site: static where content is stable, server-rendered where it is personal or fast-moving. Then we measure on real devices and mid-tier connections, not a developer laptop.

Question 21

Can it work with our existing CMS?

Accepted Answer

Usually. We integrate with headless WordPress and other content sources routinely. Keeping editors in a tool they already know is generally worth more than a technically purer rebuild nobody enjoys using.

Question 22

Why React rather than something lighter?

Accepted Answer

Because the interfaces we build tend to be long-lived and heavily stateful — dashboards, portals, configurators. React’s ecosystem and the depth of the hiring pool matter more over a five-year life than initial bundle size.

Question 23

Will we be able to hire for it?

Accepted Answer

Yes, which is part of why we choose it. A platform your team can staff and maintain independently is worth more than a marginally faster framework only a handful of people know.

Question 24

Is React the right call for a marketing site?

Accepted Answer

Often not. A content site with modest interactivity is usually better served by simpler rendering. We recommend the lighter option when it fits, because we maintain what we build.

Question 25

We already have AWS. Why add this?

Accepted Answer

Because having AWS is not the same as knowing your posture. Security Hub continuously checks your accounts against recognised benchmarks and surfaces drift, which is what turns security from an annual assertion into something you can evidence on any given day.

Question 26

Does this replace a penetration test?

Accepted Answer

No. A test is a point-in-time probe; continuous posture management catches the misconfiguration introduced three weeks later. They answer different questions and mature programmes run both.

Question 27

What do we actually receive?

Accepted Answer

A prioritised view of real findings with remediation owners and dates, reviewed with you on a regular cadence — not an unfiltered console dump. The measure of success is findings closed, not findings counted.

Question 28

Is AI visibility measurable, or guesswork?

Accepted Answer

Measurable, with caveats we state plainly. We track how your brand appears across AI answer engines for the questions your buyers actually ask, and report share and movement rather than implying precision the medium does not support.

Question 29

How is this different from SEO rank tracking?

Accepted Answer

Answer engines synthesise rather than list, so there is no single position to hold. What matters is whether you are cited, how you are characterised, and whether the citation is accurate — which is a content and structure problem as much as a ranking one.

Question 30

Can you guarantee we appear in AI answers?

Accepted Answer

No, and nobody credible can. What we can do is make you the most citable source on your own subject — structured, verifiable, current — and report honestly on where that is working and where it is not.

Question 31

What is a guardrail in practice?

Accepted Answer

An enforced boundary, not a polite instruction. Scope, tone, refusal behaviour and prohibited topics are implemented in code and tested, so the assistant cannot be talked out of them by a determined user.

Question 32

Who decides where the boundaries sit?

Accepted Answer

Your risk owners, with our input. We bring the failure modes we have seen; you own the policy. It is written down, implemented, tested, and reviewed when the business or the regulation changes.

Question 33

What happens when someone tries to break it?

Accepted Answer

It refuses, logs the attempt, and — where the context warrants — routes to a human. We test adversarial prompts as part of the evaluation suite rather than waiting to discover the gap in production.

Question 34

Why do we need evaluations at all?

Accepted Answer

Because a model update, a prompt edit or a content change can degrade an AI feature without anything appearing broken. Evaluations turn that invisible risk into a failing test, which is the only way to run AI in production responsibly.

Question 35

What do you actually measure?

Accepted Answer

Grounding — did the answer come from your content; refusal correctness — did it decline what it should decline; and regression against known-good responses. Plus latency and cost, because a correct answer nobody waits for is not a working feature.

Question 36

Who sees the results?

Accepted Answer

You do. Evaluation results and the criteria behind them are yours, reported alongside the rest of the programme. If a change makes something worse, you hear it from us with the evidence attached.

Question 37

Does AI write our production code?

Accepted Answer

AI assists; engineers decide. Every change passes human review, automated tests and the same release process as any other work. What changes is throughput on the mechanical parts — scaffolding, migrations, test coverage — not accountability.

Question 38

Does this mean lower cost or just faster delivery?

Accepted Answer

Usually faster, and better covered. The gain shows up in the work that used to get skipped under time pressure: tests, documentation, refactoring. We would rather spend the recovered time on quality than quietly bank it.

Question 39

Is our source code exposed?

Accepted Answer

No. It runs under enterprise terms where code is not used to train public models, inside our own controlled environments. For clients with stricter requirements we agree the boundary before any repository is connected.

Question 40

Does this lock us into one framework?

Accepted Answer

We use orchestration where it earns its place and plain code where it does not. The retrieval design, prompts, evaluations and data contracts are the durable assets; the orchestration layer is deliberately the most replaceable part of the stack.

Question 41

Who maintains it after launch?

Accepted Answer

Whoever you choose — us on a support arrangement, or your own team. We hand over documented pipelines, the evaluation suite and the prompts, because an AI feature nobody can safely change is a liability rather than an asset.

Question 42

How do you test something non-deterministic?

Accepted Answer

With an evaluation suite of real inputs and expected behaviours, run on every change. We measure grounding, refusal correctness and regression against known-good answers, so a prompt tweak cannot quietly degrade quality.

Question 43

Why not just fine-tune a model on our content?

Accepted Answer

Because your content changes. Retrieval means you edit a document and the answer changes immediately, with a citation back to source. Fine-tuning bakes knowledge in, is expensive to redo, and cannot show where an answer came from.

Question 44

How current are the answers?

Accepted Answer

As current as your source material. We index on a schedule appropriate to how fast your content moves — continuously for pricing and stock, daily or weekly for policy and guidance — and stale content is the thing we monitor for.

Question 45

What happens when the answer is not in our content?

Accepted Answer

It refuses and offers a route to a human. That behaviour is designed and tested, not incidental: an assistant that admits the gap is far more useful than one that fills it plausibly.

Question 46

What does MCP actually give us?

Accepted Answer

A governed way for an assistant to use your systems — a CRM lookup, a stock check, a booking — instead of only talking about them. Each tool is exposed with explicit permissions and full logging, so you can see precisely what was called and by whom.

Question 47

Is it safe to let an assistant touch live systems?

Accepted Answer

Only with boundaries, which is the point of building it this way. Read access and write access are separate decisions, anything that changes a record requires human approval in the flows we build, and every call is logged for audit.

Question 48

Do we need MCP if we already have APIs?

Accepted Answer

MCP does not replace your APIs; it sits in front of them with the permissioning, schema description and logging a language model needs to use them predictably. If you have good APIs already, most of the work is done.

Question 49

Is our data used to train the model?

Accepted Answer

No. We deploy Claude under enterprise terms where your prompts and content are not used to train public models. For clients in regulated sectors that term is usually what makes the project viable, so we confirm it in writing before any data moves.

Question 50

How do you stop it inventing answers?

Accepted Answer

We put retrieval in front of the model: it answers only from your approved content and cites the passage it used. Anything outside that scope returns a refusal rather than a guess, and automated evaluations run on every prompt or model change to catch drift.

Question 51

Could a cheaper model do the job?

Accepted Answer

Sometimes, and we will say so. Where the task is classification or extraction at volume, a smaller model is often the right call. Where the input is long, messy and consequential, correction cost outweighs licence cost — and correction cost is what decides whether the feature survives.

The stack we actually run.

Fit before fashion

Depth over breadth

No dead ends

You own it

Past the demo, into production — and into governance

Grouped by the job it does

AI & agents

Engineering & platform

Measurement

Security & assurance

The services this stack powers

Not sure which stack fits? We will tell you honestly.