Boost.ai

Conversational AI is all Boost.ai has ever sold. The company has built it since its founding in 2016, with no product underneath. The virtual agent is what an institution buys, and models do the work of understanding what a customer asked and producing the response throughout. The no code builder configures the agents rather than constituting a separate product. This is an assistant native product, not a channel platform that added models to an existing communications spine.

The oversight mechanism has a name but no published internals. The Trust Layer is presented as the control surface around the virtual agents, with guardrails intended to prevent manipulation, keep responses within approved bounds and align behavior to sector regulation. The reference deployment supports it in practice.

DNB automates a stated 20 percent of customer service and routes the remainder to people, and its executive frames the objective as an equilibrium between human and machine rather than maximal automation. The specification is missing. No confidence threshold or described trigger for handover to a person is published, and there is no approval checkpoint. Nothing accounts for what the guardrails evaluate before a response reaches a customer.

The Trust Layer is a safety control, not a validation position. No accuracy figure, containment rate or escalation rate is published for the agents, and there is no error rate or benchmark. No ISO 42001 certification is held, and no validation documentation is offered to the institution. Neither the security nor the privacy certification attests to how the models behave. An institution running virtual agents across 150 million conversations a year has a measurement surface that would support a published quality figure, and none exists.

The evidence joins named customers to a published outcome. Named customers include Nordea, DNB and Telenor. DNB, one of the largest banks in the Nordics, is the subject of a published case study reporting 20 percent of customer service automated through the platform, with a named executive quoted on the balance between human and machine handling. Independent evaluation is specific.

The company was named a Leader in the 2025 Gartner Magic Quadrant for Conversational AI Platforms, a stated placement rather than mere participation, having previously appeared in the 2022 enterprise edition. Published scale is consistent across sources at more than 600 live agents across more than 450 organizations handling more than 150 million conversations annually. Nordic Capital has held the company since 2021, which puts a party with capital at risk behind the diligence.

The company names its safety architecture and the failure modes it is built against. The Trust Layer is described as fortifying virtual agents with checks and balances, with guardrails positioned as tamperproof and resistant to jailbreak attempts. The published explanation states plainly that language models can be unpredictable and can generate biased responses. It adds that they can hallucinate and can be exploited through prompt injection. The boundary questions go unanswered. Nothing states whether conversations from one of the 450 organizations inform models serving another, or whether customer interactions are used for model improvement.

ISO 27701 certification is the distinguishing fact. It is an audited privacy information management system rather than a policy statement, held since June 2022 and scoped to the whole group including subsidiaries, with 46 of 49 controls implemented and that ratio published. The privacy policy names the regimes it addresses, which are European data protection law, California privacy law as amended, and United Kingdom data protection law.

The legal entity and its company registration number are printed. The operational documents are not public. No data processing agreement, subprocessor list or retention schedule is published outside the gated trust center, and nothing describes how long conversation transcripts are held or what is done with them.

Boost.ai publishes not only its certificates but how complete each implementation is. It has held ISO 27001 since May 2021 and ISO 27701 since June 2022, both scoped to the whole company including subsidiaries rather than to a product line. The implementation ratio for each is published, 113 of 114 security controls and 46 of 49 privacy controls. That shows where implementation falls short, which a certificate alone never does.

A trust center is operated for the full compliance overview, and alignment with OWASP guidance is stated. A third party consultancy reports a SOC 2 Type II report, but it does not appear on the company's own security page, so it remains unconfirmed.

As a technology supplier to regulated institutions, Boost.ai holds no financial license, which is the expected posture. Its positioning is built around regulated industries and its certifications are scoped accordingly, but technical certifications are not regulatory standing. No supervised regulator test, sandbox participation or program admission is public. No position under the EU AI Act is published, despite a European base and deployment inside banks and public bodies.

Bias is named as a risk the Trust Layer guardrails are built to address, and disclosure stops there. No fairness testing, differential outcome monitoring or impact assessment is published. One gap is specific and measurable. The platform operates across Norwegian, Swedish, Danish and Finnish as well as English, at minimum, and no per language accuracy or containment figure is published.

A customer served in a less represented language may be understood less reliably than one served in English. Public sector deployment sharpens this, because a citizen facing service agent that performs unevenly across languages affects access to a public entitlement rather than a commercial product.

No error rate, remediation commitment, liability position or correction path is published. The guardrail architecture is a control against a wrong answer being produced, not a route for someone who has already received one. A customer told something incorrect about an account, a payment or an entitlement by a virtual agent has no stated way to have that interaction reviewed or corrected. Nothing describes whether a transcript is retained in a form that would let the institution reconstruct what was said.

No model provider, family, version or country of processing is disclosed. The published material discusses large language models in general terms as the technology underlying modern virtual agents and identifies their risks candidly, without stating whose models the platform runs or whether they are hosted by the company or called from an external provider.

For a vendor whose whole proposition is reliability in regulated settings, and whose customers include national banks and public bodies, the question of whose infrastructure processes a citizen's or customer's words is one a buyer would expect answered.

Integration depth is undocumented beyond a listing on one enterprise software marketplace. No core banking or policy administration system is named as a supported integration, and neither is any contact center platform or data source. No integration count is published. Deployments at Nordea, DNB and Telenor make deep integration a practical certainty, so this is a disclosure gap rather than a capability one. It is still a conspicuous omission, since competing platforms publish counts in the hundreds and name the core platforms they connect to.

Private cloud deployment is offered alongside the standard hosted service, a real second deployment model. The European base and the group wide privacy certification give a buyer reasonable grounds to expect European processing, but nothing states it. No region list or residency commitment is published, and there is no single tenant description or on premises path. A bank in a jurisdiction with localization requirements would be relying on inference from the company's domicile rather than on a published position.

No pricing or tier structure is published, and there is no billing basis or indicative range. The route to a number is a demo request. The absence stands out because the company publishes precise operational figures elsewhere, including certification control ratios and conversation volumes. The reticence is specific to price rather than a general reluctance to publish numbers.

More than 450 organizations use the platform across four regulated sectors. Banking and insurance sit alongside telecommunications and the public sector, which is real breadth of buyer type. Within financial services the coverage is narrower than the headline count suggests. The installed base is concentrated in the Nordics and wider Europe, and the workload is customer service and internal employee support rather than lending, servicing or back office process execution.

Two separate competitors, each writing on its own domain, make the same architectural observation, that the platform is positioned for conversational support rather than end to end operational automation. The vendor's own material is consistent with that reading.