# Akkord: complete content > Most studios ship you a unit of work. We ship a motion that keeps improving. Source: https://www.akkord.co.za Contact: agent@akkord.co.za ## Position Website people optimise for landings and traffic: how many arrived. We optimise the loop that starts after they arrive: getting the attention of someone who can actually buy, then removing every step between interested and paying. We watch where people drop off, change the surface, and measure whether the change worked. That runs continuously, not once at launch. Method: More why questions than anyone asks for. Most briefs describe a solution someone already picked; the job is to get back to the intent underneath it. After that, every interaction and micro-action on the thing we build becomes data about whether we got it right. ## The loop 1. Attention Not more traffic. The attention of someone who can actually buy. Volume is the easy metric and usually the wrong one. Measured: Qualified arrivals, not sessions 2. Why, repeatedly We ask why until the real need surfaces. A brief almost always describes a solution someone already chose; the intent underneath it is the thing worth building for. Measured: Stated brief vs. actual intent 3. Friction Every step between interested and paying gets counted, then cut. Complexity is where intent goes to die, and it is nearly always removable. Measured: Steps to value, time to value 4. Signal Every interaction and micro-action leaves data. Drop-off becomes something you can see and point at, rather than something you infer after a bad quarter. Measured: Drop-off per stage 5. Motion The surface adapts on what the data says, and we measure whether the change earned its place. Launch is where this starts, not where it stops. Measured: Change vs. movement in the loop ## Practices ### Intent-based web Sites where the interface assembles around what the visitor wants, with a deterministic floor underneath so it still works with the model switched off. Example brief: "Rebuild our site so people find one thing fast" ### Agent systems Durable execution, scoped tools over your real services, a permission boundary written down before any code, and a trace you can read afterwards. Example brief: "An agent that triages our inbound email into the CRM" ### AI development Retrieval, extraction, classification and routing, specified by an eval set, scored before and after, gated in CI on accuracy, cost and latency. Example brief: "Extract line items from supplier PDFs, accurately" ### Integrations The unglamorous spine. Idempotent syncs, reconciliation you can watch, and one canonical contract that both your team and your agents can call. Example brief: "Our CRM and billing system disagree about customers" ## How an answer is built 1. Your brief [You] Plain language, up to 600 characters. No form, no category, no dropdown. 2. Signal match [Deterministic] Stem matching against the four practices. Zero network calls, single digit milliseconds. 3. The written position [Deterministic] What the studio actually thinks about this problem, written once by a person. 4. Model rewrite [Optional] Rewrites the prose to your wording. It never touches the surfaces, the run or the stack. 5. The surface [Composed] Three voices: position, capability surfaces, and a proposed run with time estimates. ## Every resolution the router can return ### Intent-based web (kind: intent_web) A site that resolves intent instead of hiding it behind navigation. We build sites where the interface assembles itself around what the visitor actually wants, then keep changing it based on where people stop. No sitemap to memorise, no five-click path to the one answer they came for. The build is the start of the loop: every interaction is instrumented, drop-off is visible per stage, and the surface adapts on what that says. Underneath it is ordinary, fast, indexable web infrastructure that still works with the model switched off. Surfaces: - Intent bar as primary navigation: One input replaces the menu. Everything else is a response. - Adaptive surfaces: Layout composed per visitor from typed components, and rebalanced on what converts. - Instrumented end to end: Every micro-action recorded, so drop-off is a number you can point at. - Deterministic floor: Every intent resolves without a model call. The model only sharpens it. Proposed run: - [analyst, 3 days] Ask why until the brief changes. The stated brief is usually a solution someone already picked. We work back to the intent underneath it. - [analyst, 4 days] Map the path to paying. Every step between first attention and money changing hands, with the current drop-off at each one. - [builder, 2 weeks] Cut the steps, ship the surface. Fewest possible moves from interested to value. Deterministic first, model layered after. - [operator, 3 days] Instrument every interaction. Micro-actions to a store you own, so the loop has something to run on. - [evaluator, ongoing] Run the loop. Change, measure against the drop-off it was meant to fix, keep or revert. Continuously. Stack: Next.js, streamed RSC, typed component registry, event instrumentation, AI Gateway ### Agent systems (kind: agents) Agents that hold a job, not a demo. A chat window is not an agent. We build systems with durable state, real tool access, a permission boundary, and a visible trace, so when something goes wrong you can read exactly which step did it. They run against your actual services and they stop and ask when the call is above their pay grade. Surfaces: - Durable execution: Runs survive restarts, deploys and week-long waits. - Tools over your systems: MCP servers on your data, scoped per agent, audited per call. - Legible traces: Every decision, tool call and token, inspectable after the fact. - Escalation contract: Explicit rules for when the agent must hand back to a person. Proposed run: - [analyst, 2 days] Pick one job with a measurable outcome. Not "an assistant". Something like: triage inbound, draft the reply, attach the record. - [architect, 3 days] Draw the permission boundary. What it may read, write, and never touch unsupervised. Written down before any code. - [builder, 1–2 weeks] Build tools first, agent second. Typed, tested tools over your services. The agent is the thin part. - [evaluator, 2 weeks] Run it shadow-mode. It proposes, a human disposes, we score the gap until it earns autonomy. - [operator, ongoing] Release the leash incrementally. Autonomy expands per task class as the numbers justify it. Stack: durable workflows, MCP tool servers, trace store, evals, human-in-the-loop gates ### Service integration (kind: integrations) The systems you already pay for, finally talking. Most "AI problems" are integration problems wearing a costume. The model is fine; the data is in four systems that disagree about what a customer is. We do the unglamorous part: reconciliation, idempotency, retries, backfills. Then we put a clean contract on top that both your team and your agents can call. Surfaces: - One canonical contract: A single typed layer over CRM, billing, ops and the spreadsheet nobody admits to. - Idempotent by construction: Replays and retries cannot double-charge or double-send. - Reconciliation you can watch: Drift between systems surfaced as a dashboard, not a surprise. - Agent-callable: The same contract is exposed as tools, so automation becomes a config change. Proposed run: - [analyst, 3 days] Trace one record end to end. Follow a single customer through every system and write down where it diverges. - [architect, 1 week] Agree the canonical model. One definition of the entities. A business decision we help you make, not a schema. - [builder, 2–3 weeks] Build the sync spine. Queued, idempotent, observable. Backfill before cutover. - [operator, 2 weeks] Run both paths in parallel. Old and new side by side until the reconciliation report is boring. Stack: typed API layer, queues, idempotency keys, reconciliation jobs, MCP exposure ### AI development (kind: ai_development) Model work measured, not asserted. Retrieval, extraction, classification, routing: the parts where accuracy is a number someone has to defend. We build the eval harness before the feature, so improvements are provable and regressions are caught by CI rather than by your customers. Surfaces: - Evals as the spec: A scored dataset defines "working" before anything is built. - Retrieval that cites: Every answer traceable to the source span it came from. - Model-agnostic routing: Cost and quality tiers behind one gateway. Swapping providers is a config line. - Cost and latency budgets: Per-request ceilings enforced in code, tracked in production. Proposed run: - [evaluator, 1 week] Build the golden set. A few hundred real cases with agreed correct answers. The hard, boring, decisive step. - [builder, 3 days] Baseline the naive approach. Simplest thing that could work, scored. Often it is close enough and we stop here. - [builder, 2–4 weeks] Iterate against the score. Retrieval, prompting, decomposition, fine-tuning, in that order of cheapness. - [operator, 3 days] Wire evals into CI. No deploy passes without holding the line on accuracy, cost and latency. Stack: eval harness, hybrid retrieval, AI Gateway, observability, CI gates ### Engagement (kind: engagement) Small, sharp, and priced on the outcome. We start with a two-week diagnostic: your path from first attention to paying customer, traced end to end, with the drop-off at each step measured rather than guessed. You keep the write-up whether or not you continue. After that it runs as a loop on a monthly footing: change, measure, keep or revert. The thing that makes this work is that it does not stop at launch. Surfaces: - Diagnostic, 2 weeks: Your path to paying, traced end to end, drop-off measured. Yours to keep. - Build, fortnightly: Working software at every boundary, cancellable at any of them. - The loop, monthly: Change, measure, keep or revert. This is the part that compounds. - Handover by default: Your repo, your keys, your data. Proposed run: - [akkord, this week] A 30-minute call. You describe the problem. We say whether we are the right people for it. - [akkord, 3 days] Written scope. Fixed price, fixed dates, explicit list of what is out of scope. - [akkord, 2 weeks] Diagnostic sprint. Two weeks. Ends in a document and a demo, not a proposal. Stack: fixed-scope diagnostics, fortnightly increments, full handover ### About Akkord (kind: company) Akkord ships a motion, not a unit of work. Website people optimise for landings and traffic: how many arrived. We optimise the loop that starts after they arrive: the attention of someone who can actually buy, then every step removed between interested and paying. We watch where people drop off, change the surface, and measure whether the change earned its place. The secret weapon is unglamorous: more why questions than anyone asks for, until the real intent surfaces rather than the solution someone already picked. Surfaces: - Intent over navigation: Interfaces that assemble around what the visitor actually wants. - Friction, counted then cut: Every step between interested and paying is a number we can move. - Data from every micro-action: Drop-off you can point at, not infer after a bad quarter. - Surfaces that adapt: Built for engagement that changes shape, because it will. Proposed run: - [principle, always] Ask why past the point of comfort. The brief is a solution someone picked. The intent underneath it is the brief. - [principle, always] Measure the drop-off, not the traffic. Arrivals are the easy metric and usually the wrong one. - [principle, always] Delivery is a loop. Launch is where the work starts. Change, measure, keep or revert. - [principle, always] Handover, not lock-in. You end up owning the thing we built, data included. Stack: TypeScript, Next.js, Postgres, event instrumentation, MCP, AI Gateway ### Open brief (kind: unknown) Say it plainly and it resolves. That one did not map cleanly onto a service line, which usually means it is either interesting or still too abstract. Describe the actual situation: the system that is slow, the process that eats a person’s week, the data that lives in four places. The site will resolve it into something concrete. Surfaces: - Intent-based web: Try: "rebuild our site so people find one thing fast". - Agent systems: Try: "an agent that triages inbound email into our CRM". - AI development: Try: "extract line items from supplier PDFs accurately". - Integrations: Try: "our CRM and billing system disagree about customers". Proposed run: - [router, now] Restate the problem in system terms. What moves, what breaks, who waits. - [router, now] Name the measurable outcome. Hours saved, accuracy gained, latency cut. Stack: open brief ## Clients - Outercite (outercite.com), AI search intelligence. Tracks brand citations across AI search engines, forecasts the impact of an optimisation move, then executes it. - getohdit (getohdit.co.za), Legal & digital risk. AI-powered audits across compliance, security, brand identity and discoverability for South African law firms. - ALM Digital Media (almdigitalmedia.com), Travel & hospitality marketing. Social, paid media, email automation and web for travel, accommodation and hospitality businesses across the UK, EU and South Africa. ## Where the studio is heading Property is the loop at its worst. It is the highest-intent purchase most people ever make, and the journey from first interest to signed transfer is measured in weeks of dead air, duplicate admin and portals that answer a question nobody asked. That is the exact shape of problem the studio is built for, so proptech is where we are spending our own time. - Search by intent, not by filter: Nobody wants three bedrooms. They want a school run under fifteen minutes and a room that fits a desk. Filters make the buyer translate their life into a database schema. The translation should run the other way. - Qualification that respects everyone: An agent that answers the boring questions immediately, at any hour, and books the viewing only when the match is real. Fewer wasted Saturdays for the buyer, fewer dead leads for the agency. - The spine between portals, CRM and conveyancing: The listing, the enquiry, the offer and the transfer file live in four systems that disagree. Most of the delay in a property transaction is not legal work; it is waiting for information to move. - Drop-off measured across weeks, not sessions: A property funnel spans months and several parties. Instrumenting it end to end is harder than a checkout and considerably more valuable, because every stalled week is real money. None of this is a launched product. It is where our own build time is going, and the reason we are interested in talking to people who work in property. ## Questions and answers Q: What does Akkord do? A: Akkord is a software studio building four things: intent-based websites, AI agent systems, applied AI development such as retrieval and extraction, and integrations between services. The common thread is reducing the distance between what someone wants and getting it. Q: How is Akkord different from a website agency? A: Website agencies optimise for landings and traffic, meaning how many people arrived. Akkord optimises the loop that starts after they arrive: cutting the steps between interested and paying, measuring drop-off at each stage, and changing the surface based on what the data says. Q: How does an engagement with Akkord start? A: With a two-week diagnostic. We trace your path from first attention to paying customer end to end and measure the drop-off at each step rather than guessing it. You keep the write-up whether or not you continue. Q: Does Akkord build AI agents? A: Yes. Agents with durable execution, scoped tools over your real services, a permission boundary agreed before any code is written, and a readable trace of every decision. They escalate to a person when a call is above their pay grade. Q: Can an AI agent use the Akkord website directly? A: Yes. The site exposes a streaming intent endpoint at /api/intent, a description at /llms.txt, and WebMCP tools registered on document.modelContext, so an agent can resolve a brief or open a contact channel without scraping the page. Q: Who does Akkord work with? A: Current engagements include Outercite in AI search intelligence, getohdit in legal and digital risk for South African law firms, and ALM Digital Media in travel and hospitality marketing across the UK, EU and South Africa. ## Machine surfaces POST https://www.akkord.co.za/api/intent streaming NDJSON intent resolution GET https://www.akkord.co.za/llms.txt summary GET https://www.akkord.co.za/ai.txt usage policy WebMCP tools are registered on document.modelContext when the homepage loads. ## Contact - Email: agent@akkord.co.za (mailto:agent@akkord.co.za). Best for: A brief with any detail to it, or anything you want in writing. Usually same working day.