All field notes

Xero's API is one of the cleanest in accounting software, which makes agent work straightforward and shifts the difficulty to workflow design and controls. This guide covers OAuth and rate-limit handling, the bookkeeping and advisory workflows worth automating, how practices serve many clients from one build, and the controls that keep the ledger defensible.

Zoho runs sales, support, finance, and operations for a large number of small and mid-sized businesses from one suite, which makes cross-application agents unusually valuable. This guide covers OAuth scoping across the suite, the CRM, Desk, and Books APIs, the cross-app workflows that pay first, and the controls a lean business needs without an enterprise programme.

Oracle ERP Cloud exposes good integration surfaces, which makes agent work more straightforward than in older estates and no less consequential. This guide covers REST and BI Publisher integration, role and data access security, the exception workflows worth automating first, and the audit controls a financial system demands.

SAP holds the transactions that matter, which makes it the highest-value and most carefully governed place to put an AI agent. This guide covers integration options across S/4HANA and ECC, authorisation and segregation of duties, the workflows that pay first, and the controls an SAP estate demands.

A stalled AI pilot isn't free—it's sunk spend, delayed value, and eroded trust that makes the next project harder to fund. The true cost of AI that never ships.

One agent is an engineering project; fifty agents is an operating problem. This whitepaper sets out how organisations run agents at scale: choosing between workflow and autonomous orchestration, managing state and handoffs, containing failures so one agent cannot cascade, handling identity and permissions across agents, and the shared platform and registry that keep the estate governable.

The data platform built for analytics does not serve AI. Retrieval needs curated content with ownership and freshness, agents need low-latency context serving with entitlements enforced, evaluation needs labelled reference data, and everything needs lineage an auditor can follow. This whitepaper sets out what an AI data platform contains, how it relates to the existing warehouse, and how to build it incrementally.

Compliance functions are asked to cover expanding obligations with static headcount, using sampling methods that miss most of what they are meant to detect. This whitepaper sets out where AI changes that: full-population monitoring instead of sampling, surveillance triage, policy and obligation management, regulatory change analysis, and control testing, with the evidence and human determination that regulators expect.

Security operations face two AI problems at once: using AI to handle alert volumes no team can clear, and defending AI systems that introduce attack surfaces conventional controls never anticipated. This whitepaper covers both, setting out triage and investigation support, detection engineering, automated response boundaries, and the controls that secure agents, retrieval systems, and the data behind them.

AI changes what work people do long before it changes how many people are needed, and organisations that handle the first badly never get to manage the second well. This whitepaper sets out a workforce transition practice: role redesign from task analysis, skills assessment and development, redeployment planning, honest communication, managing mixed human and digital teams, and the measures that show whether the transition is working.

Most RAG systems are evaluated by asking a few questions and reading the answers, which is why so many degrade silently after launch. This whitepaper sets out a methodology: separating retrieval and generation measurement, constructing reference sets that reflect real questions, scoring groundedness and completeness, running evaluation continuously rather than at release, and monitoring the leading indicators that move before quality does.

Every enterprise AI decision eventually reduces to the same question: buy the product built for your industry, build on a general platform, or both. This whitepaper sets out where vertical products genuinely win, where horizontal platforms do, how to compare total cost honestly including the integration and data work vendors omit, how lock-in differs between them, and the hybrid architecture most organisations arrive at.

AI agents are easy to demo and hard to run reliably. The real failure modes—scope, guardrails, evaluation, oversight—and how to ship agents that hold up.

Guardrails are what make an autonomous agent safe to run. The permission, action, and output controls that bound behavior—and how to design them well.

Field operations carry the largest controllable cost in many service businesses and the least reliable data, because the work happens away from the systems that record it. This whitepaper maps AI across dispatch and scheduling, first-time-fix improvement, technician assistance, work documentation, parts and inventory, and safety, and sets out the offline, connectivity, and adoption realities that decide whether any of it reaches the van.

M&A runs on document review under time pressure and integration work that outlasts the deal team's attention. This whitepaper sets out where AI helps across diligence, contract and obligation analysis, day-one readiness, systems and data consolidation, and synergy tracking, with the confidentiality, privilege, and clean-team constraints that govern deal work.

Procurement holds an unusual combination: enormous document volume, fragmented spend data, supplier risk obligations that keep expanding, and approval authority that cannot be delegated to software. This whitepaper maps AI across spend analysis, sourcing events, contract lifecycle, supplier risk, and purchase-to-pay operations, sets the boundary around approval and award decisions, and gives a sequence that produces measured savings and cycle-time results.

Quality management runs on records: inspections, non-conformances, corrective actions, complaints, audits, and the documentation that proves each was handled. This whitepaper sets out where AI helps, including defect detection, non-conformance triage, root cause support, complaint classification, and audit preparation, and what validation and change control require when the output touches a regulated quality record.

Sustainability reporting has moved from voluntary narrative to assured disclosure with auditor scrutiny, and the data required sits in systems and suppliers that were never built to produce it. This whitepaper sets out where AI helps: data collection and mapping across entities, supplier data gathering, emissions calculation support, assurance evidence, and disclosure drafting, all under the controls that assured reporting demands.

CRM has always promised a single view of the customer and delivered a database of what sellers were willing to type. AI changes that by capturing activity automatically, enriching records from real sources, and turning a maintenance burden into a system that gives something back. This whitepaper sets out where AI belongs across data quality, pipeline intelligence, seller productivity, forecasting, and service, and the governance customer data requires.

Not every workflow needs an AI agent—many need plain automation. When rules beat reasoning, when you need an agent, and how to choose without over-engineering.

AI spend behaves unlike any other technology cost: variable per request, sensitive to changes nobody labels as cost-relevant, and difficult to attribute from a provider invoice. This whitepaper adapts FinOps practice to AI, covering visibility and attribution, unit economics per use case, forecasting, an optimisation ladder that protects quality, chargeback models, and the monthly rhythm that keeps engineering and finance aligned.

ERP programmes are the largest, slowest, and riskiest projects most enterprises run, and AI changes their economics in specific places: data cleansing and migration, process discovery from actual system behaviour, test generation and execution, and the intelligent layer that reduces the customisation that made the last upgrade impossible. This whitepaper sets out where AI helps, where it does not, and how to keep the core clean.

Every AI vendor relationship ends, and the cost of ending it is decided years earlier by contract terms and architecture nobody examined at the time. This whitepaper sets out what an organisation must own to stay portable, where lock-in genuinely bites, how to structure contracts and architecture for exit, and how to run a migration without rebuilding from scratch.