Most chatbots deflect. Few resolve. The bots customers actually like are the ones that complete the task they came for: book the appointment, change the order, update the account, refund the charge. We build chatbots that connect to your real systems, ground every answer in approved data and hand off cleanly when they hit the boundary of what they should do.
Pixelfield is for Heads of CX, Heads of Operations and CTOs at companies where support volume is growing faster than headcount. The seniors who scope the work are the seniors who write the code. 50+ AI features in production, fastest deployment two months from kickoff to live, and a reputation for telling you when Intercom is enough.
Resolution, not just deflectionConnected to your real systemsHandoff designed as a featureFlat run-rate, no per-resolution surprises
Intercom Fin, Zendesk AI, Drift before it shut down. These are good for moderate volume, mostly-FAQ workflows where deflection to a help article is acceptable. They stop working when you need account-aware actions, custom integration, predictable flat pricing at higher volume, or a resolution rate higher than what their boilerplate model can deliver. If a platform is enough, use it. If it isn't, the next step is custom.
Loops. Confident wrong answers. 'Let me connect you with a human' followed by a blank ticket where the agent has to ask the customer to start over. A bad chatbot is worse than no chatbot. We start by acknowledging that and re-architecting around the things that actually drive CSAT: grounded answers, structured handoff, escalation that triggers on confidence rather than retries, honest 'I don't know' responses.
Booking changes, refund processing, account updates, order edits, password resets, plan upgrades. Transactional bots are most of what we build. The bot reads and writes to your real systems via OAuth-scoped permissions, irreversible actions go through human approval, every action is logged, there is a kill switch. This is the work hosted platforms can't do without significant custom development on top.
Around 10,000 conversations a month is where most platforms become the most expensive option. $0.99 per Fin resolution at 50% resolution rate is roughly $5,000 a month in resolution fees alone, before seats and add-ons. A flat-cost custom build with model routing and caching becomes the cheaper option, and the quote doesn't move when your traffic does.
Related AI cluster services
We work with your real ticket data, transcripts and knowledge base to map which queries are bot-shaped, which are deterministic and which should not be automated. Discovery includes a technical spike on the systems the bot has to talk to, because integration is where the timeline actually lives.
You receive: a conversation flow map, escalation policy, integration plan, projected resolution rate based on your real traffic, and a fixed-scope build quote.
For higher-stakes deployments we run a paid pilot on a narrow, well-scoped slice of your support volume. The pilot measures real resolution rate, handoff CSAT, hidden cleanup hours and customer reaction. We report numbers, not vibes.
We have stopped pilots when the resolution economics didn't justify the build. That's what the pilot is for.
Build runs as a small senior team led by Michal Vavra, with AI and integration engineers embedded with your CX team. Each release ships behind feature flags with conversation regression tests, output validation, escalation logging, drift detection and cost telemetry in place from week one.
You receive: a working bot in staging, then production. Knowledge grounding is treated as a first-class concern. The bot cites sources or admits uncertainty.
Live deployment with monitoring on resolution by topic, handoff rate, CSAT delta, cost per resolution and knowledge freshness. Every two weeks we review where the bot is failing and feed those cases back as KB updates, prompt changes or escalation tweaks.
Industry benchmark: 90 to 120 days from launch to first measurable ROI lift. We size the retainer to that horizon. Take it in-house whenever you're ready.
Direct answers to the questions Heads of CX and CTOs ask us on every <strong>scoping call</strong>.
Deflection means the bot kept the conversation away from a human, often by sending a link. Resolution means the customer's task was actually completed. Industry benchmarks across 220M+ AI interactions land around 45% fully AI-resolved without human handoff for general support, rising to 70-90% for narrow, well-scoped use cases (Comm100, Botpress, Gartner). We size the proposal to your traffic mix and tell you honestly which band you're likely to land in.
If your queries are mostly FAQ-shaped and your volume is moderate, those platforms are usually the right choice and we'll say so. Custom development pays off when you need account-aware actions, deeper integration, flat predictable pricing at higher volume, or resolution rates above what the platform's boilerplate can deliver. At 10,000 conversations a month, per-resolution pricing on a hosted bot can easily run $5,000+ in fees alone. Our flat retainer becomes the cheaper option.
Layered mitigations: retrieval grounded in approved sources only, output validation against typed contracts, forced tool calls before any state-changing claim, confidence thresholds that escalate instead of guess, and faithfulness scoring tracked per release. The bot is configured to say 'I don't have that information, let me get someone who does' rather than improvise. We measure the actual hallucination rate and ship mitigations until it's acceptable for your use case.
Multi-signal escalation: explicit request, low confidence, frustration sentiment, repeated rephrasing, complexity rules, topic-based routing. The agent receives a structured handoff payload: full transcript, AI-generated 2-3 sentence summary, intent, sentiment flag, customer/account state, suggested next action. The agent opens with 'I see you were asking about order #123' rather than 'how can I help you?' That single design choice is the largest CSAT lever in a chatbot system.
Discovery is fixed-fee from £2K. Production builds typically start at £8K and run six to ten weeks. Run rate is flat monthly: hosting, model inference (with routing and caching), retainer for monitoring and KB sync. We quote both numbers in the proposal so there are no per-resolution surprises later.
Yes. Transactional bots are most of what we build. The bot can read and write to your CRM, helpdesk, billing system, booking engine and inventory via OAuth-scoped permissions. Irreversible actions go through human approval or an explicit policy check. Every action is logged. There is a kill switch.
Industry data lands first measurable lift between 90 and 120 days after launch (data accumulation, prompt tuning, knowledge base maturation). Full ROI payback typically inside 6 to 14 months in well-scoped deployments. We design the retainer and the review cadence around that horizon. Roughly 95% of GenAI pilots fail to show P&L impact (MIT 2025); the difference is narrow scoping, real measurement and disciplined iteration, all of which we treat as part of the build.
