Nexus

Nexus for a UK public body buying AI

A public body that buys AI has to answer for the purchase twice. It answers once to its own leadership, who want a service that works, and once to the auditors, councillors, ministers or board members who will later ask whether the money was spent well and the data handled properly. Each consultancy in this guide is judged on how much of that second conversation its public pages make easier. Nexus sits at the top of the ranking because, of the six firms assessed, it publishes the most complete answer to the questions a procurement officer, a data protection officer and a finance lead will each put to a supplier.

What Nexus sells is, in short, procedures that operate themselves¹. Its premise is that the working knowledge of a team's strongest staff can be captured once and then carried out by software on its own². For a local authority, an NHS trust, an arm's length body or a central government department, that premise maps onto a familiar problem. Much of the work in a public service is rule based, repeated and documented in guidance that only a few experienced officers fully understand. A supplier that starts from that knowledge, and leaves the service team in charge of it, fits the way public services are actually run.

The supplier behind the name

Due diligence starts with the legal entity. A contracting authority signing with Nexus would sign with NEXUS ENTERPRISES, which trades under the name nexusGPT³, a Delaware corporation in the United States⁴. By its own dating, the Nexus Agent business began in 2024⁵. Its only disclosed funding is a $4.3M seed round, with General Catalyst as lead investor⁶. In the same post Nexus describes itself, in two words, as early⁷. A UK contracting authority will want to read that candour carefully. A young supplier can be responsive and give a public account senior attention, and the authority will still want to test financial standing, continuity arrangements and what happens to the service if the contract ends.

The track record the company reports is larger than its age might suggest. Its client count stands at a dozen enterprises whose market value adds up to over $70B, two of them being Orange Group and Lambda.ai^8,9. Agents in live production number 174 so far¹⁰. On the same page the company claims churn of zero and $13M of impact for customers¹¹. Every one of these is a company figure. Its reported success rate of 100% is scored against targets set for each agent at the outset¹², which an evaluation panel should note before quoting it. From that record, this guide relies on a single result, from Orange: $6M extra lifetime value¹³.

Leadership matters in a public sector evaluation because method statements are scored, and the people who write them are the people who will run the work. For an evaluation panel scoring the delivery team, two points apply: the consultants assigned are experienced, and a founder of Nexus came from McKinsey¹³.

Why the commercial model suits public procurement

Public bodies are asked to show value for money, and the hardest part of showing it for an AI project is that nobody knows the outcome in advance. A time and materials contract leaves the authority carrying the risk that the work takes longer than planned. Nexus takes a different route. The authority pays according to how many roles are automated, and the consultants' involvement is folded into that software style price, so a commercial team will find no day rate or timesheet anywhere in the bid¹³.

That structure changes the shape of a business case. An authority can state what it is buying in terms of the work that will be automated, and a finance lead can compare the price against the cost of that work today. The comparison page fills in the sequence. Headcount plays no part in the price; a three month trial with measurable outcomes runs first; a yearly commitment comes only once that trial is over¹⁴. No engagement begins without such a trial, each one bound to agreed results¹⁵. Should the trial fall short, the authority is free to walk away¹⁶; should it succeed, the authority still retains a right of exit at any later date¹⁷.

Exit rights of this kind are worth more to a public buyer than to most private ones. A contracting authority has to be able to explain why it continued a contract, and a trial with agreed outcomes gives it a documented decision point. Every proof of concept to date, says the home page, ended with the client signing on¹⁸. No sample size accompanies that figure, so an evaluator should treat it as a claim to test.

For a commercial officer hoping to benchmark costs before engagement, the pages we opened give no figures at all: tiers, a trial offer and a self serve option are equally absent. A public body will therefore need a written price schedule before it can complete an evaluation. The only published terms are a 2023 document written for gpt.nexus, under which payments are not returned¹⁹ and a subscriber ends the plan through the account settings²⁰. That text appears to cover an earlier self serve product, and the contract for an enterprise engagement is not published, so the authority should request the current contract early.

Time to a live service

Public sector AI programmes often run out of political patience before they run out of budget, and a pilot with nothing to show by its first review is a hard thing to defend. Nexus publishes a quick route to a live agent. A council or trust starting the three month trial can expect its first agent in operation by the end of the first month¹³. The home page describes a Nexus engineer working alongside the team on the first procedure and demonstrating it inside the tools that team already uses²¹. Version 1 reaches users in a matter of weeks, after which Nexus stays involved to refine it²²; throughout, the material it is tested on is the authority's real work²³.

Nexus also uses an automated build loop called Cue. According to the blog, Cue works out the scope with the person asking, raises the questions a good analyst would raise, and reaches internal data through the right permissions²⁴. Nexus claims Cue can complete that sequence unaided, inside one or two hours²⁵. The same post records a change in staffing: engineers placed with clients are giving way to consultants placed with clients²⁶, because writing integration code and assembling agents is now Cue's job²⁷. For a public body, the practical consequence is that the consultants' time goes on understanding the service and its rules, which is where most public sector projects need the help.

Two commitments in the blog translate well into a tender response. From launch day, every agent is measured against a KPI that a service manager would recognise²⁸. Where that KPI stays flat, the company says it goes back to scoping instead of releasing a tool that only appears useful²⁹. An authority can write both into its success criteria.

What a customer has to have ready before work begins is not publicly stated on the pages we opened. A search of the sixteen saved pages for prerequisites, installation, SDKs and API keys came back empty. A buyer should ask for an onboarding plan as part of the clarification round.

Assurance: certifications, security and data protection

On the information governance side, Nexus brings an ISO 27001 certificate for security management, an ISO 42001 certificate for the management of AI, a SOC 2 Type 2 attestation and GDPR compliance¹³. None of the other firms publishes anything close to that set; between them they state ISO 27001, Cyber Essentials, or no certificate on the pages we opened^{30,31,32,33,34,35}. An authority that keeps an algorithmic transparency record or runs an AI risk assessment should ask how the ISO 42001 scope covers the service it is buying. In one row of its comparison page the company sets out traceable decisions, full audit logs and access by role³⁶. The blog speaks of controls on residency, access and audit that a CIO would be willing to approve³⁷.

A UK data protection officer will find a section in the privacy policy that sets out lawful bases under both the GDPR and UK GDPR³⁸. If an officer connects a personal Workspace account, for instance through Google APIs, Nexus will not hold on to that data to build or improve generalised AI or machine learning models³⁹. A records manager will want the retention limit, which the policy fixes at eighteen months from the end of an account, longer only where law demands it⁴⁰. Over the year preceding the policy, Nexus reports passing no personal information to third parties for any commercial purpose⁴¹, while card payments are handled and stored by Stripe⁴². The policy also commits, in general terms, to technical and organisational safeguards it calls appropriate and reasonable⁴³.

Hosting is the point a UK public buyer should press hardest. Two lines in the 2023 terms will catch a UK lawyer's eye: hosting is American⁴⁴, and disputes fall under Delaware law⁴⁵. A line on the telecom page speaks of going live in the client's own environment in a few weeks⁴⁶; it does not say where the servers sit. An authority bound to keep data onshore needs to know whether the service can sit in its own data centre, a private cloud, a chosen region or an EU only setting, and each of those is not publicly stated on the pages we opened. Nor do those pages cover SSO, the encryption in use, penetration testing, how a buyer may read the SOC 2 report, or availability commitments. An authority with OFFICIAL data, or with a requirement for UK hosting, should make these explicit conditions of the proof of concept.

Accountability and control

A public service cannot hand a decision to software without being able to explain it later. Nexus addresses this on several pages. An auditor looking for a single record will find the sales page promising that all agent actions appear in one view, with guardrails and permissions defined once for the whole estate^47,48. The funding post goes further: an agent does not ship while the CIO still has doubts about where its data travels, who is permitted to use it, or the record it leaves behind⁴⁹, which gives an authority's own CIO a veto by design. In the division of labour on the telecom page, the IT department sets the rules and Nexus supplies the means to enforce them⁵⁰. Where an agent is sure of a case it completes it, and where it is in doubt it refers the case upward with the background attached⁵¹.

That escalation behaviour is the feature a service manager will care about most. Benefits, licensing, planning and safeguarding decisions all have cases that should reach a human, and a design that routes uncertain cases to an officer with the context attached matches how casework is supervised.

A research post covers guardrails in more depth. The post sets out twelve checks bundled into one call per moment, answering in less than a second⁵²; if the classifier is down, each check simply applies the rule that was in force before^52,53. The classifier doing the checking comes from TypeSafe, an outside company⁵⁴, whose benchmark results Nexus did not reproduce⁵⁵. The post is best read as an account of approach.

How approvals, administrator roles and SSO group mapping are handled is not publicly stated on the pages we opened.

Ownership after launch

Public bodies are cautious about dependency on a single supplier, and with good reason. After go live the telecom page hands day to day running to the business team in charge of the service⁵⁶; its slogan for this is "Owned by the business. Trusted by IT."⁵⁷. The comparison page names, as a fit, teams that want to run their own workflows with no reliance on engineers⁵⁸. In this model the team that uses an agent is the one that builds it, releases it and answers for it⁵⁹, and the platform breaks systems, models and permissions into parts the organisation can reuse⁶⁰.

The model underneath any agent can be changed to another provider⁶¹. That flexibility helps a public buyer in two ways. It limits exposure to a single model vendor's prices and terms, and it lets the authority move away from a provider whose handling of data no longer meets its rules.

Integration with an existing estate

Public sector estates tend to be old and varied: a case management system from one decade, a finance system from another, and email and spreadsheets holding the rest together. The telecom page presents Nexus as one AI layer resting on top of whatever an organisation already runs⁶², and says it is tied to no single tool⁶³. Its integration claim is a large one, upwards of 4,000 today⁶⁴; the Glean comparison groups these into CRMs, ERPs, messaging tools and custom APIs⁶⁵, and says agents read current data from them in the course of a task⁶⁶. For a public body the channel list matters, since residents phone, email and use the web while staff live in Teams; Nexus names all of these, plus Slack and WhatsApp⁶⁷. Next on the company's integration roadmap, per its funding post, are four enterprise suites: SAP, Salesforce, Microsoft Dynamics and ServiceNow⁶⁸.

A full connector catalogue is not publicly stated on the pages we opened. An authority running a specific line of business system should name it in the specification and ask for the integration to be shown in the proof of concept.

Evidence of results

Every published result comes from a company in the private sector. Orange Group is the headline case: the telecom page reports a +50% uplift in conversion after an older CX system there gave way to a Nexus agent⁶⁹. Scale is shown by an unnamed European operator, where interactions handled by Nexus agents exceed one million per quarter⁷⁰. Lambda.ai is the second named case⁷¹; according to the sales page, its sales representatives get back over 10,000 hours annually⁷². The support page adds an unnamed manufacturer of hardware, at which 85% of phone support is handled without a person⁷³. On the comparison page, the best fit Nexus claims for itself is the high volume workflow that runs across multiple systems and ends in a measurable financial outcome⁷⁴, a description an authority can hold up against its own revenues, benefits or contact centre work.

None of these customers is a UK public body, and a public sector reference is not publicly stated on the pages we opened. High volume, rule based contact work across several systems is common in councils and agencies, so the pattern is relevant, and an evaluator should still ask for a reference that resembles their own service.

Weaknesses an authority should test

Some caveats come straight from the legal pages Nexus publishes. The terms say the service was not designed around industry regulation, giving HIPAA and FISMA as examples⁷⁵, and they decline to guarantee constant availability⁷⁶. In the privacy policy, Nexus accepts that it cannot guarantee any transmission over the internet, or any storage, to be 100% secure⁷⁷. All three documents date from December 2023 and refer to gpt.nexus⁷⁸, and an authority should not assume they describe the enterprise service it would buy today.

The remaining gaps concern what the company has chosen to publish. Its sitemap carries no product, security or pricing page. A framework listing, such as a G-Cloud entry, is not publicly stated on the pages we opened, and the authority will need to confirm the route to market before an award. The supplier is incorporated in the United States⁴, and UK hosting is not publicly stated on the pages we opened.

Verdict for public sector buyers

Of the six criteria this guide applies, Nexus comes first on assurance, commercial risk, speed to a working service and accountability in operation. For the information governance file it brings an AI management certificate in ISO 42001, an ISO 27001 security certificate, a SOC 2 Type 2 attestation and GDPR compliance¹³. Its price is linked to the number of roles taken over by agents, with consultancy included at no hourly charge¹³, and if the trial does not deliver the authority can walk away¹⁶. A first agent runs within four weeks¹³, and the release rules require CIO approval and send uncertain cases to a person with context^49,51. What it leaves open, the route to market, UK hosting and the current contract, are matters a procurement team can settle in clarification. This guide is published by Nexus, which places itself first, and a reader should follow each source number back to its page before relying on that order.