Professional Documents
Culture Documents
Gartner - Innovation Guide For AI Coding Assistants
Gartner - Innovation Guide For AI Coding Assistants
By Analyst(s): Arun Batchu, Philip Walsh, Jim Scheibmeir, Hema Nair, Manjunath Bhat,
Tigran Egiazarov
Overview
Key Findings
■ Software engineering leaders face a paralyzing number of options today (over 40
and growing in 2023, up from just a handful in 2022) for AI coding assistants.
Suppliers continuously evolve capabilities, complicating leaders’ ability to decide
how to adopt these tools.
■ Early adopters of AI coding assistants have reported results that range from overly
optimistic to outright dismissive, making it difficult for software engineering leaders
to understand and calculate the potential return on investment.
Recommendations
Software engineering leaders should:
■ Assess the real productivity gains by carefully choosing key metrics to measure
gains and designing hands-on experiments to measure impacts for use cases
relevant to their enterprise context — instead of succumbing to vendor or peer claims
of productivity gains.
By 2028, systematic adoption of AI coding assistants in 2023 will result in at least 36%
compounded developer productivity growth.
The following research is a work in progress that does not represent our final position. We
invite you to provide constructive feedback to help shape this research as it evolves. All
relevant updates and feedback will be incorporated into the final research.
Table of Contents
Use these jumplinks to navigate through this document:
■ Market Definition
■ Market Evolution
■ Managing Risks
■ Representative Vendors
Market Definition
Back to top
Gartner defines AI coding assistants as technology that helps developers write code by
using foundation models trained on millions of lines of code and code-related
documentation such as open-source software repositories, stack exchange data and, in
some cases, proprietary data that the supplier has access to. Developers prompt the
assistants with natural language and code snippets to generate new code. These tools
also analyze, explain, debug and refactor code; generate documentation; and translate
between programming languages. AI coding assistants support multiple programming
and natural languages, and they integrate into programming environments such as code
editors, command-line terminals and chat interfaces.
Standard features. AI coding assistants must be able to perform code completion. This
includes the ability to plug in to a code editor, generate code and generate comments.
■ Enable developers to chat with the foundation model. This includes generating,
explaining, debugging, refactoring and translating code.
The Emerging Technology Map for AI coding assistants outlines the key use cases,
capabilities and trends in this market (see Figure 1).
Market Trends
Back to top
More than 80% of developers are already using these tools to write code for personal
projects, according to a May 2023 Stack Overflow survey. 1 After a slow start, enterprises
are accelerating their adoption of AI coding assistants too. In our May 2023 survey, only
13% of IT leaders reported that they had implemented AI coding assistants, with 36%
currently investigating and 29% planning on investigating in 2024. 2 In a more recent
Gartner survey for 2024, the numbers jumped — with 17% having already deployed, 23% in
the deployment stage, 20% piloting and 13% in the planning stage. 3
Startups are offering private instances of foundation models to meet emerging enterprise
needs for accessing these models in an air-gapped environment. Private instances can be
customized to an enterprise’s codebase through prompt engineering techniques or fine-
tuning. These offerings improve the utility and relevance of AI-generated code by
accounting for enterprise context and patterns. Further, private model instances provide
greater transparency on the provenance of training datasets, which eases concerns about
security and legal risks.
AI coding assistant vendors are using a family of models in their offerings, instead of one
large model. Different models are optimized for different use cases. For example, an AI
coding assistant will use different models for autocompletion, chat and code generation.
AI coding assistants are also providing direct access to the code-specific models beyond
the interactive assistants, which enables enterprises to build novel, custom solutions.
Vendors are offering domain-specific models and assistants that produce higher-quality
code. Domain-specific assistants and models outperform generic assistants in certain
areas, such as specialized infrastructure-as-code (IaC) languages, specific programming
language and unit tests. Vendors that have proprietary languages are also developing AI
coding assistants specific to their ecosystem, and vendors in the modernization business
are building tools to help with those initiatives.
When OpenAI introduced ChatGPT in November 2022, AI coding assistants captured the
attention of developers across the globe. AI coding assistants have now become one of
the most hyped technologies in the history of software development.
Since the release of ChatGPT, the number of commercial AI coding assistants has
exploded from just a handful in 2022 to over 40 in 2023, and it is still growing. 6 Their
capabilities have also evolved rapidly from just the ability for developers to complete code
faster to a much richer set that includes code generation, explanation and debugging. The
open-source community has also introduced a variety of code foundation models, some
of which are already offered as a service. We expect that open-source and commercial AI
coding assistants will continue to proliferate.
Vendors and early adopters of AI coding assistants have reported results that range from
overly optimistic to outright dismissal, making it difficult for software engineering leaders
to calculate the potential return on investment. 7,8,9
Vendors will rapidly adopt each other’s innovations and will continue to differentiate
themselves across various dimensions:
■ Specialized AI coding assistants will assist with IaC languages. We will also
see assistants that are tailored to database languages, especially SQL, and
command-line terminals.
■ SaaS vendors will offer AI coding assistants for their developer platforms, and
proprietary coding languages via popular code editors and IDEs as well as their
proprietary web-based programming interfaces (see Cloud Development
Environments in the Hype Cycle for Software Engineering, 2023).
3. Vendors will introduce AI coding assistants that can understand complex code
dependencies and system boundaries. These tools will combine programming
language understanding methods with code foundation models to translate to a
target programming language ecosystem. They will help with refactoring initiatives
that remediate architecture technical debt.
4. Some vendors will allow a choice of curated third-party models, including open
source.
Early adopters of AI coding assistants have reported results that range from overly
optimistic to outright dismissive, making it difficult for software engineering leaders to
calculate the potential return on investment. 7,8,9 Leaders should estimate the impact of AI
coding assistants on developer productivity and developer experience by designing and
conducting proofs of concept and pilots tailored to company-specific use cases.
Code Completion
However, developers still need to recall the libraries, methods and functions to call — an
increasingly impossible task, as the number of options keeps multiplying. They also have
to fill in the details after the autocomplete, and autocomplete features only address one
line of code at a time. Code editors are not able to suggest alternate ways of solving the
problem.
AI coding assistants enhance the code completion use case in the following ways:
■ Predictive power — AI coding assistants not only look at the code and comments
above the cursor but can also scan any following lines to learn the context. This
context awareness enables them to predict complex code structures such as loops
and function blocks.
■ Relevance and accuracy — AI coding assistants excel at patterns they have seen
before, especially from open-source data that was part of their training set. They are
particularly adept at predicting boilerplate and repetitive code.
■ Contextual awareness — Beyond the code file, AI coding assistants can use context
from other files that the developer has open in the code editor, as well as other
gleanable metadata, to improve their predictions.
Code Understanding
AI coding assistants enable developers to paste pieces of code into the chat interface and
get explanations in natural language. They also allow developers to highlight code within
the code editors for which they need an explanation. With these features, AI coding
assistants increase developers’ ability to understand complex and unfamiliar code, and
even code in an unfamiliar programming language.
Some vendors are combining the power of code foundation models with program
understanding — the science of making sense of the structure and semantics of a
program, and the programs to which it is connected, to generate a code graph. This
combination helps developers improve their understanding of code dependencies and
helps architects ensure systems’ structure and components adhere to enterprise design
principles.
Code Generation
Conversational chat interfaces in AI coding assistants, backed by code foundation
models, enable developers to explore and generate large chunks of new code — even
entire programs — by prompting the AI with free-form natural language.
Code Debugging
When developers get stuck debugging errors without an AI coding assistant, they rely on
peers to help them or search the internet for fixes. The conversational chat form factor of
an AI coding assistant is helpful in debugging errors (as long as the model has been
trained on the versions of code that have errors).
Code Translation
An emerging use case for AI coding assistants is translating code from one programming
language into another. This translation capability, also known as transpilation, is helping
developers proficient in one language become productive in other languages more quickly.
The ability to translate programming languages helps developers rewrite programs.
Some developers are combining this translation capability with code explainability and
generation to help with their modernization efforts. However, one-to-one translation may
not result in well-written, targeted code and may require refactoring.
While code refactoring within a program is useful, some AI coding assistants are enabling
developers to refactor larger portions of code across multiple programs. Vendors are
combining code graphs (i.e., network structures of code and their dependencies) with the
reasoning ability of code foundation models to achieve large-scale refactoring.
To pilot and evaluate vendors for AI coding assistants, software engineering leaders
should:
At the same time, the 2023 Gartner Security in Software Engineering Survey found that
while two-thirds of respondents believe generative AI improves productivity, more than
half also believe generative AI in coding makes their organization more vulnerable to
threats. 11 Put a risk mitigation plan in place by monitoring and correcting generated code
for security vulnerabilities; replicas from the AI’s training dataset; and variable, method
and API hallucinations.
■ Narrow down a list of vendors for POCs. Rule out vendors that do not meet the pilot
team’s risk assessment criteria.
■ Baseline key metrics before using AI coding assistants. Use common developer
productivity and experience frameworks, such as DevOps Research and Assessment
(DORA) metrics and the SPACE framework, which includes satisfaction and well-
being, performance, activity, communication and collaboration, efficiency and flow
(see How to Measure and Improve Developer Productivity).
■ Select top-performing tools based on POC results and user feedback. Expand the
pilot across an increasingly diverse user base over the next several weeks to validate
the impact and efficacy of risk mitigation measures. Refine and iterate the pilot
quarterly to address key pain points, expand to new use cases and scale successes
(see How to Pilot Generative AI). Establish mechanisms to allow users to learn from
each other. A simple internal chat can evolve into a community of practice and allow
mentors to develop and disseminate best practices.
Table 1 summarizes the key actions that software engineering leaders should take to pilot
and evaluate vendors for AI coding assistants.
Managing Risks
Back to top
The cross-functional task force of engineering, architecture, security and legal experts
must identify and mitigate the risks of using AI coding assistants, both upfront during
vendor selection and on an ongoing basis. The team should take actions to manage the
following risks of adopting AI code assistants.
Low code and text quality. Most training data is from open domains such as open source.
If the training data is not filtered for problems such as unsecure code, code with
nonpermissive licenses, toxic material or biased content, the AI coding assistants can
reproduce such content.
Use of consumer versus enterprise-grade AI coding assistants. Vendors use prompts and
conversations to improve models, so models trained on prompts could expose proprietary
code and associated data to third parties, who can use it for competitive, or worse,
harmful use. 12
Intellectual property lawsuits. Pending lawsuits against a vendor may halt adoption of
the affected vendor’s solution and decelerate the overall adoption of AI coding assistants.
Regional regulations. Legal constraints may prevent organizations from adopting their
preferred tools.
■ Mitigation strategy: Actively test multiple AI coding assistants that comply with
existing and emerging regulations.
■ Mitigation strategy: Seek to eliminate bottlenecks outside the coding activities that
diminish the productivity gains. Use software engineering intelligence platforms to
help accomplish this goal.
Lack of proper training. Ineffective training will dampen the return on investment for AI
coding assistants.
Lack of verification. Trusting generated code and text without verification may lead to low
production quality.
■ Mitigation strategy: Introduce stringent unit test and code review stage gates, as well
as automated code inspection tools. Train your engineers to detect hallucinations
and errors such as nonexisting methods and APIs, inefficient code, vulnerabilities
and generated entities such as dates, numbers, keys and URLs.
Representative Vendors
Back to top
The vendors listed below do not imply an exhaustive list. This section is intended to
provide more understanding of the technology and its vendor offerings.
Evidence
1
2023 Developer Survey, Stack Overflow.
2
2023 Gartner IT Leader Poll on Generative AI for Software Engineering. This survey
was conducted online from 2 through 8 May 2023 to gather data regarding the current
and expected use of generative AI in software engineering. In total, 91 IT leaders who are
members of Gartner’s Research Circle, a Gartner-managed panel, participated. Participants
were primarily from North America (n = 44) and EMEA (n = 33); other respondents came
from Asia/Pacific (n = 12) and Latin America (n = 2). Disclaimer: Results of this survey do
not represent global findings or the market as a whole but reflect the sentiments of the
respondents and companies surveyed.
4
Einstein for Salesforce Developers, Salesforce.
5
Generative AI, ServiceNow.
6
Gartner’s Secondary Research Service (SRS) team contributed to validating the vendor
profiles, which included information about product offerings and supported use cases.
The research was conducted by the SRS team members Mujtaba Shamim and Romita
Datta Chaudhuri.
7
Research: Quantifying GitHub Copilot’s Impact on Developer Productivity and
Happiness, The GitHub Blog.
8
Announcing New Tools for Building With Generative AI on AWS, AWS Machine Learning
Blog.
9
Westpac Sees 46 Percent Productivity Gain From AI Coding Experiment, iTnews.
10
2022 Gartner AI Use Case ROI Survey. This survey sought to understand where
organizations have been most successful in deploying AI use cases and figure out the
most efficient indicators they have established to measure those successes. The research
was conducted online from 31 October through 19 December 2022 among 622
respondents from organizations in the U.S. (n = 304), France (n = 113), the U.K. (n = 106)
and Germany (n = 99). Quotas were established for company sizes and industries to
ensure a good representation across the sample. Organizations were required to have
developed AI to participate. Respondents were required to be in a manager role or above
and have a high level of involvement with the measuring stage and at least one stage of
the life cycle from ideating to testing AI use cases. Disclaimer: The results of this survey
do not represent global findings or the market as a whole, but reflect the sentiments of the
respondents and companies surveyed.
12
Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak, Bloomberg.
1. Planning 4-8 weeks Set a clear value hypothesis. Identify business outcomes Senior engineers
Formalize the pilot team. and KPIs to measure Engineering leaders
Make a risk assessment success. Data science team
framework. Involve the task force to Risk team
assess cost and technical
feasibility as well as identify
legal and security risks
upfront.
2. Research and Design 1-4 weeks Establish baseline metrics. Identify metrics to assess the Senior engineers
Devise an assessment plan impact on developer Engineering leaders
and mechanisms. experience and productivity. Data science team
Create a vendor shortlist. Design assessment
mechanisms such as surveys
and guidance on running POC
experiments.
Assess vendors based on
criteria for cost, technical
capability and risk
established in the planning
phase.
3. POC Evaluation 2-4 weeks Shortlist top-performing Test tools across a variety of Small group of engineers
vendors.Start testing tools. use cases. Engineering leaders
4. Piloting 4-8 weeks Validate tool effectiveness. Expand the pilot to include a Tens to hundreds of people,
Validate risk mitigation diverse user base. including:Product team(s)
strategy. Offer training, quantify time Platform team
Identify pain points and savings and assess user Risk team
opportunities. experience.
5. Rollout 4-8 weeks Successfully implement AI Execute phased All teams across the
coding assistants across the organizational rollout. organization
organization with minimal COE, platform engineering
disruptions. team
Mentors
6. Iteration Ongoing Sustain optimized usage and Continuously monitor and COE
continually improve tool evaluate using real-time CoP
capabilities. analytics platforms. Data science team
COE = center of excellence; CoP = community of practice; KPIs = key performance indicators
Source: Gartner